The reviewed record of science sign in
Pith

arxiv: 2607.06560 · v1 · pith:U7RMFRXY · submitted 2026-07-07 · cs.CV

Vision as Unified Multimodal Generation

Xiaoyang Han , Jianhua Li , Kewang Deng , Zukai Chen , Xuanke Shi , Sihan Wang , Boxuan Li , Linyan Wang
show 9 more authors
Siyi Xie Xin You Jinsheng Quan Zhongang Cai Haiwen Diao Ziwei Liu Lei Yang Dahua Lin Quan Wang
This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-08 01:38 UTCglm-5.2pith:U7RMFRXYrecord.jsonopen to challenge →

classification cs.CV
keywords unified multimodal generationcomputer visioninstruction-responsedense predictionsegmentationmulti-view geometryvision foundation model
0
0 comments X

The pith

One model, no task-specific heads, matches vision specialists

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that the full spectrum of computer vision — from object detection and OCR to depth estimation, segmentation, and multi-view 3D geometry — can be cast as unified multimodal generation within a single pretrained model that natively handles both text and image synthesis. The central object is SenseNova-Vision, a model built by fine-tuning an existing unified multimodal model (a UMM) on a large-scale corpus of computer-vision annotations converted into instruction-response pairs. The key mechanism is a data protocol that maps each vision task into the model's native output spaces: symbolic records like bounding boxes and keypoints become text generation, dense spatial predictions like depth maps and segmentation masks become image generation, and compositional tasks use mixed text-and-image outputs. No task-specific prediction heads or architectural changes are introduced. The authors show that this single model matches or approaches leading task-specialized systems across structured visual understanding, dense geometric prediction, segmentation, and multi-view visual geometry, while retaining the base model's general multimodal abilities. The paper also releases the training corpus and model publicly.

Core claim

The paper's central claim is that heterogeneous computer vision tasks — spanning detection, OCR, keypoints, segmentation, depth, surface normals, point maps, and camera pose estimation — can all be expressed through the native text and image generation spaces of a single unified multimodal model, without task-specific architectures, and that the resulting model achieves competitive performance against specialist systems across all four major vision task families. The discovery is that the GPT-style consolidation paradigm from NLP transfers to computer vision when dense spatial outputs are represented as images in the model's native image-generation space and symbolic outputs as text, with a

What carries the argument

The load-bearing mechanism is the data conversion protocol that maps each vision task into one of three native generation channels: text generation for symbolic records (bounding boxes, OCR strings, keypoints, camera parameters), conditional image generation for dense spatial predictions (depth maps as grayscale, surface normals as RGB, segmentation masks as color-coded images, point maps as XYZ-encoded RGB), and mixed text-and-image outputs for compositional tasks like grounded conversation generation segmentation. Camera pose estimation uses 2,001 reserved vocabulary tokens encoding quantized quaternion rotations, translation directions, and scales. Training uses standard cross-entropy for

If this is right

  • If the formulation is correct, new vision tasks could be added by writing new instruction templates and converting annotations, rather than designing new architectures, lowering the cost of extending perception capabilities.
  • Language-defined task variants that recombine capabilities across domains — such as using detection coordinates to drive segmentation, or OCR to drive text-region masks — could emerge from joint training without explicit supervision for each variant.
  • The approach provides a concrete path for absorbing computer vision supervision into general-purpose foundation models, potentially allowing vision capabilities to scale with language model and image generation model improvements rather than requiring separate specialist pipelines.
  • The staged convergence pattern observed — with dense geometric tasks converging fastest and multi-view geometry slowest — suggests that different vision abilities have different inherent difficulty under the generation formulation, which could inform training curricula for future unified models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The VAE latent space acts as an information bottleneck for dense geometric outputs; tasks requiring fine geometric precision (multi-view reconstruction, camera pose) show the largest gaps versus specialists, which may reflect a fundamental precision ceiling imposed by image-space encoding rather than a data limitation that more training can overcome.
  • If the generation-based pathway fundamentally limits geometric precision, the approach may face a performance ceiling on tasks requiring sub-pixel or high-bit-depth geometric accuracy, regardless of training data scale — suggesting that some vision tasks may resist full unification under this formulation.
  • The qualitative probes showing language-controlled mask protocols and cross-domain task recombination suggest the model is learning shared spatial-language correspondences, but the imperfect color fidelity and boundary precision in free-form probes indicate these are emergent but not yet reliable capabilities.

Load-bearing premise

The paper assumes that the VAE latent space of the base model can faithfully encode and decode dense geometric outputs — depth maps, surface normals, point maps, and segmentation masks — with sufficient precision for benchmark-compatible evaluation. The entire approach depends on image generation as the output channel for dense predictions, but no quantitative analysis is provided of the information loss introduced by VAE encoding and decoding of geometric signals.

What would settle it

If the VAE bottleneck introduces systematic geometric precision loss that cannot be recovered by training data scale, then the approach would face a hard performance ceiling on tasks requiring fine geometric accuracy — multi-view reconstruction and camera pose estimation would remain perpetually below specialist systems regardless of further training. The multi-view geometry results (e.g., ETH3D F1: 72.2 vs. 80.9 for the leading specialist) are consistent with this hypothesis.

read the original abstract

We formulate computer vision as unified multimodal generation, where heterogeneous visual tasks are expressed in the native text and image generation spaces of a unified multimodal model, without task-specific architectures. Under this formulation, SenseNova-Vision uses natural-language instructions and optional visual prompts to specify tasks, target regions or views, and decoding conventions, and generates responses as text for symbolic outputs, images for dense spatial predictions, or mixed text-and-image outputs for compositional tasks. To support large-scale training, we convert diverse computer vision annotations into instruction-response examples compatible with these generation spaces, resulting in the SenseNova-Vision Corpus, a computer-vision instruction-response corpus spanning text, image, and mixed targets. Starting from an off-the-shelf pretrained unified multimodal model, SenseNova-Vision is trained primarily on this corpus, with auxiliary multimodal data used as a capability-preserving mixture, and requires no task-specific prediction heads or architectural modifications. The resulting model covers a broad range of vision tasks, including detection, OCR, keypoint estimation, segmentation, depth estimation, surface normal prediction, point maps, and camera pose estimation, while supporting language-defined variants that combine category, color, region, and other visual cues. Experiments show that a single unified model can match leading task-specialized systems across structured visual understanding, dense geometric prediction, segmentation, and multi-view visual geometry. These results suggest unified multimodal generation as a scalable route for integrating computer vision capabilities into general-purpose foundation models. The model and corpus are publicly available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 7 minor

Summary. This paper proposes formulating computer vision tasks as unified multimodal generation within a single UMM (unified multimodal model), without task-specific prediction heads. The authors convert diverse vision annotations (detection, depth, normals, segmentation, multi-view geometry, etc.) into instruction-response examples with text, image, or mixed outputs, constructing the SenseNova-Vision Corpus (SN-VC, 50M+ examples). Starting from Bagel-7B, they train SenseNova-Vision and evaluate across four task families: structured visual understanding, dense geometric prediction, segmentation, and multi-view visual geometry. Results show competitive performance against task-specialized systems on several benchmarks (e.g., NYUv2 depth δ1=98.1, COCO-Common detection F1=56.6), with gaps on multi-view geometry (ETH3D F1=72.2 vs. VGGT's 80.9). The model and corpus are publicly released.

Significance. The paper makes a strong systems contribution: it demonstrates that a single UMM can produce benchmark-decodable outputs across an unusually broad range of vision tasks without task-specific heads, which is a meaningful step for the field. The public release of both model weights and a 50M-example corpus is a significant community resource. The data protocol for converting heterogeneous annotations into text/image/mixed generation targets is clearly documented and reproducible. The convergence analysis (Figure 6) and the qualitative language-defined task variant probes (Sec. 5.6.3–5.6.4) add useful empirical insight. However, the central paradigm-level claim—that the unified generation formulation itself is responsible for the results—lacks direct ablation support, which limits the strength of the conceptual contribution relative to the engineering contribution.

major comments (3)
  1. The paper's central conceptual claim is that unified multimodal generation (text + image generation without task-specific heads) is a viable paradigm for computer vision. However, no ablation isolates the contribution of the unified formulation from data scale and model capacity. Specifically, there is no comparison of the same Bagel-7B base with task-specific prediction heads (e.g., a depth decoder, a mask decoder) trained on the same SN-VC corpus versus the unified generation approach. Without this, the competitive results could be attributable to the 50M+ training examples and 7B model capacity rather than the formulation itself. This is load-bearing for the paradigm-level thesis stated in the abstract and Sec. 1. A controlled comparison on at least one task family (e.g., dense geometry or segmentation) would substantially strengthen the claim.
  2. The abstract states the model 'matches leading task-specialized systems across structured visual understanding, dense geometric prediction, segmentation, and multi-view visual geometry.' This is an overstatement for multi-view visual geometry: Table 4 shows ETH3D F1 of 72.2 vs. 80.9 for VGGT and 76.6 for Depth Anything 3, and CO3Dv2 AUC@30 of 80.1 vs. 91.8 for Depth Anything 3. The paper itself acknowledges this gap in Sec. 5.4 ('a performance gap remains on several metrics'). The abstract should be revised to accurately reflect the competitive-but-not-matching position on multi-view geometry, consistent with the honest assessment in the body.
  3. Section 4 and Appendix A.2 describe using MoGe-2 [177] to generate pseudo-labels for depth and normal training data, including for datasets like SA-1B (4.4M frames) and Objects365 (1.3M frames). Since MoGe-2 is also a baseline in Table 2 (marked with asterisk for re-evaluation), there is a risk of circularity: the model is trained on pseudo-labels from a system it is then compared against. The paper should explicitly acknowledge this relationship and discuss whether the depth/normal comparisons against MoGe-2 are informative given the training data pipeline. This does not affect comparisons against other baselines (DepthAnything, Lotus-2, etc.) but should be disclosed.
minor comments (7)
  1. Table 2: The asterisk footnote for MoGe-2 states it was 're-evaluated to ensure a direct and consistent comparison,' but does not clarify whether MoGe-2 was re-evaluated with the same decoding pipeline used for SenseNova-Vision's image-to-depth conversion. Please clarify.
  2. Table 5b: The footnote marks Vision Banana depth scores with ‡ for absolute-depth protocol vs. SenseNova-Vision's affine-invariant evaluation, making these non-direct comparisons. Consider moving these entries to a separate reference table or more prominently flagging the non-comparability.
  3. Figure 6: The y-axis is labeled 'Normalized Metric' but the normalization procedure is not stated. Please specify whether each metric is normalized by its final-step value or by a fixed reference.
  4. Section 5.6.4: The free-form language-to-mask probes (Figures 9–12) are qualitative only. If any quantitative metric (even on a small held-out set) could be reported for these composed task variants, it would strengthen the claim that unified training enables cross-task recombination.
  5. Appendix Table 17: The reserved token range <-1000> to <1000> for camera pose encoding is described, but the quantization resolution (1 part in 2001) and its impact on pose precision are not analyzed. A brief note on whether this resolution is sufficient for the reported camera pose accuracy would be helpful.
  6. The paper uses 'parameter-free' phrasing in places (e.g., Sec. 1: 'without task-specific prediction heads or architectural modifications'). While technically accurate, the reserved camera-pose tokens (Sec. 4, Appendix Table 17) and the 200-color palette (Appendix A.3) are task-specific design choices. Consider clarifying what 'no task-specific' encompasses.
  7. Sec. 5.5: SenseNova-Vision obtains 79.0 on MMVP vs. 83.3 for Bagel, and 0.85 on GenEval vs. 0.82 for Bagel. The MMVP drop (4.3 points) is not discussed. A brief note on whether this degradation is expected or acceptable would contextualize the capability-retention claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee identifies three major points: (1) the absence of an ablation isolating the unified formulation from data scale and model capacity, (2) an overstatement in the abstract regarding multi-view geometry results, and (3) potential circularity from using MoGe-2 as both a pseudo-label source and a baseline. We address each below and commit to revisions for points 2 and 3. For point 1, we provide a substantive response explaining why a fully controlled head-vs-generation ablation is not straightforwardly implementable within our framework, while acknowledging the referee's concern and proposing a partial remedy.

read point-by-point responses

  1. Referee: No ablation isolates the contribution of the unified formulation from data scale and model capacity. Specifically, no comparison of the same Bagel-7B base with task-specific prediction heads trained on the same SN-VC corpus versus the unified generation approach. This is load-bearing for the paradigm-level thesis.

    Authors: We agree that a controlled ablation comparing task-specific heads against unified generation on the same base and data would strengthen the paradigm-level claim. However, implementing this comparison fairly is non-trivial and we want to be transparent about why. The core architectural premise of our work is that a UMM (Bagel-7B) uses a single shared transformer backbone with a VAE-based image decoder and a text decoder—there is no separate vision backbone onto which one can trivially bolt a depth decoder or mask decoder without fundamentally altering the architecture. Adding task-specific heads to Bagel-7B would require either (a) introducing new decoder modules that are architecturally foreign to the UMM framework, which conflates the formulation comparison with an architecture comparison, or (b) training a different base model from scratch, which changes the base model and thus does not isolate the formulation variable. That said, we acknowledge the referee's concern that our results could be partially attributable to data scale and model capacity rather than the formulation itself. We propose a partial remedy: we will add an ablation on one task family (dense geometry) where we train Bagel-7B with a standard task-specific fine-tuning setup (e.g., attaching a lightweight depth prediction head to the visual encoder) on the same depth data from SN-VC, and compare against our image-generation approach. This will not be a perfect isolation of the formulation variable, but it will provide the most informative comparison we can honestly construct. We will also add explicit discussion of this limitation in the revision, noting that fully disentangling formulation from architecture remains an open challenge for the paradigm-level thesis. revision: partial

  2. Referee: The abstract states the model 'matches leading task-specialized systems across structured visual understanding, dense geometric prediction, segmentation, and multi-view visual geometry.' This is an overstatement for multi-view visual geometry: Table 4 shows ETH3D F1 of 72.2 vs. 80.9 for VGGT and 76.6 for Depth Anything 3, and CO3Dv2 AUC@30 of 80.1 vs. 91.8 for Depth Anything 3. The abstract should be revised to accurately reflect the competitive-but-not-matching position on multi-view geometry.

    Authors: The referee is correct. The abstract's use of 'matches' overstates the multi-view geometry results. Table 4 clearly shows gaps on ETH3D (F1 72.2 vs. VGGT's 80.9) and CO3Dv2 (AUC@30 80.1 vs. Depth Anything 3's 91.8), and Section 5.4 already acknowledges this gap honestly. We will revise the abstract to state that the model achieves competitive performance across the four task families, with leading or near-leading results on structured visual understanding, dense geometric prediction, and segmentation, while approaching but not yet matching the strongest specialist systems on multi-view visual geometry. This will align the abstract with the body's assessment. revision: yes

  3. Referee: Section 4 and Appendix A.2 describe using MoGe-2 to generate pseudo-labels for depth and normal training data, including for datasets like SA-1B and Objects365. Since MoGe-2 is also a baseline in Table 2 (marked with asterisk for re-evaluation), there is a risk of circularity. The paper should explicitly acknowledge this relationship and discuss whether the depth/normal comparisons against MoGe-2 are informative given the training data pipeline.

    Authors: The referee raises a valid point about potential circularity. We will add an explicit acknowledgment in both Section 4 and Appendix A.2 that MoGe-2 is used as a pseudo-label generator for depth and normal training data (for datasets without dense annotations, such as SA-1B and Objects365) and is also included as a re-evaluated baseline in Table 2. We will add a note to Table 2 clarifying this relationship. Regarding whether the comparison is informative: the MoGe-2 pseudo-labels are used only to densify incomplete supervision for a subset of training data (datasets with sparse or no depth/normal annotations), while the majority of dense geometry training data comes from synthetic datasets with ground-truth annotations (e.g., Hypersim, IRS, TartanAir, SceneNet RGB-D). Furthermore, our model is not trained to reproduce MoGe-2's outputs on benchmark evaluation sets—those sets are excluded from training. Therefore, the comparison is not fully circular, but we agree it should be disclosed so readers can interpret it appropriately. Comparisons against other baselines (DepthAnything, Lotus-2, Marigold, etc.) are unaffected by this relationship. revision: yes

Circularity Check

0 steps flagged

No significant circularity found; central claim validated on external benchmarks with standard metrics

full rationale

The paper proposes a formulation (casting vision tasks as unified multimodal generation) and validates it empirically on external benchmarks (COCO, NYUv2, ETH3D, RefCOCO, ScanNet, etc.) with standard metrics. The formulation itself is a design choice, not a derivation from prior results, so there is no derivation chain that could reduce to its own inputs. The data construction pipeline uses external tools for pseudo-labeling: MoGe-2 [177] for depth/normal labels, LingBot-Depth [159] for sparse depth completion, and Rex-Omni [75] for detection/OCR data. While MoGe-2 also appears as a baseline in Table 2, this is not circular: the model is trained on MoGe-2 pseudo-labels for non-benchmark images and then evaluated on held-out benchmark splits (the paper explicitly states benchmark images are excluded from training). This is standard knowledge distillation, not circularity. Several self-citations exist (Bagel [33] as the base model, ConsistCompose [146] for color-instance binding format, SenseNova-SI [16] as complementary work), but none are load-bearing for the central claim — the claim stands or falls on the external benchmark results, which are independently reproducible. The score of 2 reflects minor self-citation presence (Bagel as base model, some shared authors on ConsistCompose) that does not undermine the independent empirical validation.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 1 invented entities

See axiom_ledger above.

free parameters (6)
  • Training mixture weights = See Fig. 5b: 26.46% structured, 10.29% dense geometry, 15.21% segmentation, 19.31% multi-view 3D, 28.74% original multim
    The proportion of each task family and auxiliary data in the training mixture is hand-tuned, not derived from principle.
  • SigLIP2 input resolution (980px) = 980
    Chosen to preserve high-resolution conditioning for segmentation; higher than Bagel's default. Not derived from an optimization.
  • Max views per training sample = 10
    Memory-constrained choice for multi-view training; affects multi-view geometry performance.
  • Color palette (200 anchors) = 200
    Number of color anchors for instance segmentation mask encoding, chosen by greedy farthest-point sampling in RGB cube.
  • Learning rate (2.5e-5) = 2.5e-5
    Standard SFT learning rate, not tuned specifically for this task.
  • EMA ratio (0.995) = 0.995
    Inherited from Bagel; not independently justified.
axioms (4)
  • domain assumption Dense geometric outputs (depth, normals, point maps, masks) can be faithfully encoded and decoded through the VAE latent space of the base UMM with benchmark-compatible precision.
    This is the foundational assumption of the entire approach. The paper demonstrates it empirically but does not prove it analytically. It enters in §3.1 (Dense geometric prediction) and §4 (Training).
  • domain assumption Text-serialized coordinates (normalized to 3 decimal places) provide sufficient precision for benchmark evaluation of detection, keypoints, and camera pose.
    Used throughout structured visual understanding (§3.1) and camera pose estimation (§3.1, Appendix A.4). The rounding to 3 decimal places is stated in Appendix A.1.
  • domain assumption Joint mixed-task training with interleaved text and image objectives does not cause catastrophic interference between task families.
    The mixed-task fine-tuning strategy (§4) assumes that CE loss for text and rectified-flow loss for images can coexist in the same optimization without destructive interference. The convergence analysis (§5.6.1) provides partial support but no formal guarantee.
  • domain assumption Pseudo-labels generated by MoGe-2 and LingBot-Depth are sufficiently accurate for training depth, normal, and point-map prediction.
    Used in §3.2 and Appendix A.2/A.4 for data construction. The accuracy of these pseudo-labels directly bounds the achievable quality of the trained model.
invented entities (1)
  • Reserved camera-pose tokens (<-1000> to <1000>, <frame>, <quat>, <offset>, <scale>) independent evidence
    purpose: Encoding camera pose parameters as discrete text tokens for multi-view geometry tasks.
    These are not new physical entities but a tokenization scheme. The paper provides falsifiable evaluation: camera pose accuracy is measured on Re10K and CO3Dv2 (Table 4), showing competitive but not leading performance.

pith-pipeline@v1.1.0-glm · 53189 in / 4231 out tokens · 340492 ms · 2026-07-08T01:38:09.973581+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

231 extracted references · 231 canonical work pages · 54 internal anchors

  1. [1]

    Dataone-synthetic-v1.0-sample, 2025

    51WORLD. Dataone-synthetic-v1.0-sample, 2025. URL https://huggingface.co/datasets/51WORLD/ DataOne-synthetic-v1.0-sample. Accessed: 2026-06-19

    work page 2025

  2. [2]

    Ttpla: An aerial-image dataset for detection and segmentation of transmission towers and power lines

    Rabab Abdelfattah, Xiaofeng Wang, and Song Wang. Ttpla: An aerial-image dataset for detection and segmentation of transmission towers and power lines. In Proceedings of the Asian Conference on Computer Vision, 2020

    work page 2020

  3. [3]

    Matting human datasets

    AISegment. Matting human datasets. GitHub repository, 2019. URL https://github.com/aisegmentcn/matting_ human_datasets. Accessed: 2026-06-18

    work page 2019

  4. [4]

    Flamingo: A visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Bi´nkowski...

    work page 2022

  5. [5]

    IDDA: A large-scale multi-domain dataset for autonomous driving

    Emanuele Alberti, Antonio Tavera, Carlo Masone, and Barbara Caputo. IDDA: A large-scale multi-domain dataset for autonomous driving. IEEE Robotics and Automation Letters, 5(4):5526–5533, 2020. doi: 10.1109/LRA.2020.3009075

    work page doi:10.1109/lra.2020.3009075 2020

  6. [6]

    Open-world text-specified object counting,

    Niki Amini-Naieni, Kiana Amini-Naieni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting,

    work page

  7. [7]

    arXiv preprint arXiv:2306.01851

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    2d human pose estimation: New benchmark and state of the art analysis

    Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3686–3693, 2014. doi: 10.1109/CVPR.2014.471

    work page doi:10.1109/cvpr.2014.471 2014

  9. [9]

    SceneScript: Reconstructing Scenes With An Autoregressive Structured Language Model

    Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, Jakob Engel, Edward Miller, Richard Newcombe, and Vasileios Balntas. Scenescript: Reconstructing scenes with an autoregressive structured language model, 2024. arXiv preprint arXiv:2403.13064

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9535–9545, 2024

    work page 2024

  11. [11]

    Qwen3-VL Technical Report

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    Zerowaste dataset: Towards deformable object segmentation in cluttered scenes

    Dina Bashkirova, Mohamed Abdelfattah, Ziliang Zhu, James Akl, Fadi Alladkani, Ping Hu, Vitaly Ablavsky, Berk Calli, Sarah Adel Bargal, and Kate Saenko. Zerowaste dataset: Towards deformable object segmentation in cluttered scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19134–19143, June 2022

    work page 2022

  13. [13]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott ...

    work page 1901

  14. [14]

    CDLA: A chinese document layout analysis dataset

    buptlihang. CDLA: A chinese document layout analysis dataset. GitHub repository, 2021. URL https://github.com/ buptlihang/CDLA. GitHub repository, accessed June 18, 2026

    work page 2021

  15. [15]

    The 2019 davis challenge on vos: Unsupervised multi-object segmentation

    Sergi Caelles, Jordi Pont-Tuset, Federico Perazzi, Alberto Montes, Kevis-Kokitsi Maninis, and Luc Van Gool. The 2019 davis challenge on vos: Unsupervised multi-object segmentation. arXiv, 2019

    work page 2019

  16. [16]

    Rethinking object detection in retail stores

    Yuanqiang Cai, Longyin Wen, Libo Zhang, Dawei Du, and Weiqiang Wang. Rethinking object detection in retail stores. In The 35th AAAI Conference on Artificial Intelligence (AAAI 2021), 2021. 18

    work page 2021

  17. [17]

    Scaling spatial intelligence with multimodal foundation models

    Zhongang Cai, Ruisi Wang, Chenyang Gu, Fanyi Pu, Junxiang Xu, Yubo Wang, Wanqi Yin, Zhitao Yang, Chen Wei, Tongxi Zhou, et al. Scaling spatial intelligence with multimodal foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7879–7890, 2026

    work page 2026

  18. [19]

    Chameleon: Mixed-Modal Early-Fusion Foundation Models

    Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. doi: 10.48550/arXiv.2405.09818

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.09818 2024

  19. [20]

    Industrial-site-safety-detection-v1-dataset

    Chappieut. Industrial-site-safety-detection-v1-dataset. Hugging Face dataset, 2026. URL https://huggingface.co/ datasets/Chappieut/Industrial-Site-Safety-Detection-v1-DATASET. MIT License. Accessed: 2026-06-18

    work page 2026

  20. [21]

    Shikra: Unleashing Multimodal LLM's Referential Dialogue Magic

    Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal LLM’s referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  21. [22]

    Fleet, and Geoffrey Hinton

    Ting Chen, Saurabh Saxena, Lala Li, David J. Fleet, and Geoffrey Hinton. Pix2Seq: A language modeling framework for object detection. In International Conference on Learning Representations, 2022

    work page 2022

  22. [23]

    Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David Fleet, and Geoffrey E. Hinton. A unified sequence interface for vision tasks. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 31333–31346. Curran Associates, Inc., 2022

    work page 2022

  23. [24]

    Large-scale structure from motion with semantic constraints of aerial images

    Yu Chen, Yao Wang, Peng Lu, Yisong Chen, and Guoping Wang. Large-scale structure from motion with semantic constraints of aerial images. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 347–359. Springer, 2018

    work page 2018

  24. [25]

    Intern vl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Zhong Muyan, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Intern vl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24185–24198, 2023

    work page 2024

  25. [26]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024

    work page 2024

  26. [27]

    Schwing, Alexander Kirillov, and Rohit Girdhar

    Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1290–1299, 2022

    work page 2022

  27. [28]

    Domain adaptation for traffic density estimation

    Luca Ciampi., Carlos Santiago., Joao Paulo Costeira., Claudio Gennaro., and Giuseppe Amato. Domain adaptation for traffic density estimation. In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - V olume5: VISAPP,, pages 185–195. INSTICC, SciTePress, 2021. ISBN 978-989-758-488...

    work page doi:10.5220/0010303401850195 2021

  28. [29]

    The cityscapes dataset

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Scharw¨achter, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset. In CVPR Workshop on The Future of Datasets in Vision, 2015

    work page 2015

  29. [30]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    work page 2016

  30. [31]

    ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly- annotated 3d reconstructions of indoor scenes, 2017. arXiv preprint arXiv:1702.04405

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  31. [32]

    Objaverse: A Universe of Annotated 3D Objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects, 2022. arXiv preprint arXiv:2212.08051

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  32. [33]

    Smith, Hanna Hajishirzi, Ross Girshick, Ali Farhadi, and Aniruddha Kembhavi

    Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, Jiasen Lu, Taira Anderson, Erin Bransom, Kiana Ehsani, Huong Ngo, YenSung Chen, Ajay Patel, Mark Yatskar, Christopher Callison-Burch, Andrew Head, Rose Hendrix, Favyen Bastani, Eli VanderBilt, Nathan Lambert...

    work page 2025

  33. [34]

    Emerging Properties in Unified Multimodal Pretraining

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan. Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [35]

    Coconut: Modernizing coco segmentation

    Xueqing Deng, Qihang Yu, Peng Wang, Xiaohui Shen, and Liang-Chieh Chen. Coconut: Modernizing coco segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    work page 2024

  35. [36]

    SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture

    Haiwen Diao, Penghao Wu, Hanming Deng, Jiahao Wang, Shihao Bai, et al. SenseNova-U1: Unifying multimodal understand- ing and generation with NEO-unify architecture. arXiv preprint arXiv:2605.12500, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  36. [37]

    Object detection in aerial images: A large-scale benchmark and challenges

    Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Object detection in aerial images: A large-scale benchmark and challenges. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7778–7796, Nov 2022. ISSN 1939-3539. doi: 10.1109/ tpami.20...

    work page arXiv 2022

  37. [38]

    Visdrone-det2019: The vision meets drone object detection in image challenge results

    Dawei Du, Pengfei Zhu, Longyin Wen, Xiao Bian, Haibin Lin, Qinghua Hu, Tao Peng, Jiayu Zheng, Xinyao Wang, Yue Zhang, Liefeng Bo, Hailin Shi, Rui Zhu, Aashish Kumar, Aijin Li, Almaz Zinollayev, Anuar Askergaliyev, Arne Schumann, Binjie Mao, Byeongwon Lee, Chang Liu, Changrui Chen, Chunhong Pan, Chunlei Huo, Da Yu, DeChun Cong, Dening Zeng, Dheeraj Reddy P...

    work page 2019

  38. [39]

    MoRe: Motion-aware feed-forward 4d reconstruction transformer

    Juntong Fang, Zequn Chen, Weiqi Zhang, Donglin Di, Xuancheng Zhang, Chengmin Yang, and Yu-Shen Liu. MoRe: Motion-aware feed-forward 4d reconstruction transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28914–28924, June 2026

    work page 2026

  39. [40]

    Mvtec d2s: Densely segmented supermarket dataset

    Patrick Follmann, Tobias B¨ottger, Philipp H¨artinger, Rebecca K¨onig, and Markus Ulrich. Mvtec d2s: Densely segmented supermarket dataset. In Computer Vision – ECCV 2018: 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part X, page 581–597, Berlin, Heidelberg, 2018. Springer-Verlag. ISBN 978-3-030-01248-9. doi: 10.1007/978-3...

    work page doi:10.1007/978-3-030-01249-6 2018

  40. [41]

    Panoptic nuScenes: A Large-Scale Benchmark for LiDAR Panoptic Segmentation and Tracking

    Whye Kit Fong, Rohit Mohan, Juana Valeria Hurtado, Lubing Zhou, Holger Caesar, Oscar Beijbom, and Abhinav Valada. Panoptic nuscenes: A large-scale benchmark for lidar panoptic segmentation and tracking. arXiv preprint arXiv:2109.03805, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  41. [42]

    nuscenes revisited: Progress and challenges in autonomous driving

    Whye Kit Fong, Venice Erin Liong, Kok Seang Tan, and Holger Caesar. nuscenes revisited: Progress and challenges in autonomous driving. ArXiv, abs/2512.02448, 2025

    work page arXiv 2025

  42. [43]

    Image Generators are Generalist Vision Learners

    Valentin Gabeur, Shangbang Long, Songyou Peng, Paul V oigtlaender, Shuyang Sun, Yanan Bao, Karen Truong, Zhicheng Wang, Wenlei Zhou, Jonathan T Barron, Kyle Genova, Nithish Kannen, Sherry Ben, Yandong Li, Mandy Guo, Suhas Yogin, Yiming Gu, Huizhong Chen, Oliver Wang, Saining Xie, Howard Zhou, Kaiming He, Thomas Funkhouser, Jean-Baptiste Alayrac, and Radu ...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  43. [44]

    Virtual worlds as proxy for multi-object tracking analysis,

    Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis,

    work page

  44. [45]

    arXiv preprint arXiv:1605.06457

    work page internal anchor Pith review Pith/arXiv arXiv

  45. [46]

    Visual bridge: Universal visual perception representations generating

    Yilin Gao, Shuguang Dou, Junzhou Li, Zhiheng Yu, Yin Li, Dongsheng Jiang, and Shugong Xu. Visual bridge: Universal visual perception representations generating. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 21234–21242, 2026. doi: 10.1609/aaai.v40i25.39268

    work page doi:10.1609/aaai.v40i25.39268 2026

  46. [47]

    Mapillary Vistas Validation for Fine-Grained Traffic Signs: A Benchmark Revealing Vision-Language Model Limitations

    Sparsh Garg and Abhishek Aich. Mapillary vistas validation for fine-grained traffic signs: A benchmark revealing vision- language model limitations, 2025. arXiv preprint arXiv:2508.02047

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  47. [48]

    A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images

    Yuying Ge, Ruimao Zhang, Lingyun Wu, Xiaogang Wang, Xiaoou Tang, and Ping Luo. A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images. CVPR, 2019

    work page 2019

  48. [49]

    Vision meets robotics: The KITTI dataset

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. The International Journal of Robotics Research, 32(11):1231–1237, 2013. 20

    work page 2013

  49. [50]

    GenEval: An object-focused framework for evaluating text-to- image alignment

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. GenEval: An object-focused framework for evaluating text-to- image alignment. In Advances in Neural Information Processing Systems, volume 36, pages 52132–52152, 2023

    work page 2023

  50. [51]

    Precise detection in densely packed scenes

    Eran Goldman, Roei Herzig, Aviv Eisenschtat, Jacob Goldberger, and Tal Hassner. Precise detection in densely packed scenes. In Proc. Conf. Comput. Vision Pattern Recognition (CVPR), 2019

    work page 2019

  51. [52]

    Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing

    Ke Gong, Xiaodan Liang, Dongyu Zhang, Xiaohui Shen, and Liang Lin. Look into person: Self-supervised structure-sensitive learning and a new benchmark for human parsing. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017

    work page 2017

  52. [53]

    Instance-level human parsing via part grouping network

    Ke Gong, Xiaodan Liang, Yicheng Li, Yulong Chen, Ming Yang, and Liang Lin. Instance-level human parsing via part grouping network. In Proceedings of the European conference on computer vision (ECCV), pages 770–785, 2018

    work page 2018

  53. [54]

    LVIS: A dataset for large vocabulary instance segmentation

    Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019

    work page 2019

  54. [55]

    Synthetic data for text localisation in natural images

    Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In IEEE Conference on Computer Vision and Pattern Recognition, 2016

    work page 2016

  55. [56]

    Minneapple: A benchmark dataset for apple detection and segmentation.IEEE Robotics and Automation Letters, 5(2):852–858, Apr 2020

    Nicolai Hani, Pravakar Roy, and V olkan Isler. Minneapple: A benchmark dataset for apple detection and segmentation.IEEE Robotics and Automation Letters, 5(2):852–858, Apr 2020. ISSN 2377-3774. doi: 10.1109/lra.2020.2965061

    work page doi:10.1109/lra.2020.2965061 2020

  56. [57]

    Lotus-2: Advancing Geometric Dense Prediction with Powerful Image Generative Model

    Jing He, Haodong Li, Mingzhi Sheng, and Ying-Cong Chen. Lotus-2: Advancing geometric dense prediction with powerful image generative model. arXiv preprint arXiv:2512.01030, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  57. [58]

    Lotus: Diffusion-based visual foundation model for high-quality dense prediction

    Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Zhang, Bingbing Liu, and Ying-Cong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. In International Conference on Learning Representations, 2025

    work page 2025

  58. [59]

    PartImageNet: A Large, High-Quality Dataset of Parts

    Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. arXiv preprint arXiv:2112.00933, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  59. [60]

    Partimagenet: A large, high-quality dataset of parts

    Ju He, Shuai Yang, Shaobo Yang, Hengyi Zhao, Yuxin Chen, Xiaodong Li, Xingyu Qi, Yu Shen, Wei Zhang, Jing Dong, et al. Partimagenet: A large, high-quality dataset of parts. In Proceedings of the European Conference on Computer Vision (ECCV), pages 128–145, 2022

    work page 2022

  60. [61]

    Mask R-CNN

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Girshick. Mask R-CNN. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2961–2969, 2017

    work page 2017

  61. [62]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll ´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000–16009, 2022

    work page 2022

  62. [63]

    Icpr2018 contest on robust reading for multi-type web images

    Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. Icpr2018 contest on robust reading for multi-type web images. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 7–12, 2018. doi: 10.1109/ICPR.2018.8546143

    work page doi:10.1109/icpr.2018.8546143 2018

  63. [64]

    Scaling out-of-distribution detection for real-world settings

    Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. ICML, 2022

    work page 2022

  64. [65]

    Lvis fruits and vegetables

    Henning Heyen. Lvis fruits and vegetables. Hugging Face dataset, June 2026. URLhttps://huggingface.co/datasets/ henningheyen/LVIS_Fruits_And_Vegetables. Accessed: 2026-06-18

    work page 2026

  65. [66]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, volume 33, pages 6840–6851, 2020

    work page 2020

  66. [67]

    TrashCan: A Semantically-Segmented Dataset towards Visual Detection of Marine Debris

    Jungseok Hong, Michael Fulton, and Junaed Sattar. Trashcan: A semantically-segmented dataset towards visual detection of marine debris, 2020. arXiv preprint arXiv:2007.08097

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  67. [68]

    Meng-Ru Hsieh, Yen-Liang Lin, and Winston H. Hsu. Drone-based object counting by spatially regularized regional proposal networks. In The IEEE International Conference on Computer Vision (ICCV). IEEE, 2017

    work page 2017

  68. [69]

    G 2VLM: Geometry grounded vision language model with unified 3d reconstruction and spatial reasoning

    Wenbo Hu, Jingli Lin, Yilin Long, Yunlong Ran, Lihan Jiang, Yifan Wang, Chenming Zhu, Runsen Xu, Tai Wang, and Jiangmiao Pang. G 2VLM: Geometry grounded vision language model with unified 3d reconstruction and spatial reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9535–9546, June 2026. 21

    work page 2026

  69. [70]

    Deepmvs: Learning multi-view stereopsis

    Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    work page 2018

  70. [71]

    Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V . Jawahar. Icdar2019 competition on scanned receipt ocr and information extraction. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1516–1520, 2019. doi: 10.1109/ICDAR.2019.00244

    work page doi:10.1109/icdar.2019.00244 2019

  71. [72]

    Semantic Segmentation of Underwater Imagery: Dataset and Benchmark

    Md Jahidul Islam, Chelsey Edge, Yuyang Xiao, Peigen Luo, Muntaqim Mehtaz, Christopher Morse, Sadman Sakib Enan, and Junaed Sattar. Semantic Segmentation of Underwater Imagery: Dataset and Benchmark. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE/RSJ, 2020

    work page 2020

  72. [73]

    Object co-skeletonization with co-segmentation

    Koteswar Rao Jerripothula, Jianfei Cai, Jiangbo Lu, and Junsong Yuan. Object co-skeletonization with co-segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3881–3889, 2017. doi: 10.1109/CVPR. 2017.413

    work page doi:10.1109/cvpr 2017

  73. [74]

    Fashionpedia: Ontology, segmentation, and an attribute localization dataset

    Menglin Jia, Mengyun Shi, Mikhail Sirotenko, Yin Cui, Claire Cardie, Bharath Hariharan, Hartwig Adam, and Serge Belongie. Fashionpedia: Ontology, segmentation, and an attribute localization dataset. In Computer Vision – ECCV 2020, pages 316–332. Springer, 2020. doi: 10.1007/978-3-030-58452-8 19

    work page doi:10.1007/978-3-030-58452-8 2020

  74. [75]

    Megasynth: Scaling up 3d scene reconstruction with synthesized data,

    Hanwen Jiang, Zexiang Xu, Desai Xie, Ziwen Chen, Haian Jin, Fujun Luan, Zhixin Shu, Kai Zhang, Sai Bi, Xin Sun, Jiuxiang Gu, Qixing Huang, Georgios Pavlakos, and Hao Tan. Megasynth: Scaling up 3d scene reconstruction with synthesized data,

    work page

  75. [76]

    arXiv preprint arXiv:2412.14166

    work page internal anchor Pith review Pith/arXiv arXiv

  76. [77]

    ChatRex: Taming Multimodal LLM for Joint Perception and Understanding

    Qing Jiang, Gen Luo, Yuqin Yang, Yuda Xiong, Yihao Chen, Zhaoyang Zeng, Tianhe Ren, and Lei Zhang. Chatrex: Taming multimodal llm for joint perception and understanding, 2024. arXiv preprint arXiv:2411.18363

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  77. [78]

    Detect anything via next point prediction

    Qing Jiang, Junan Huo, Xingyu Chen, Yuda Xiong, Zhaoyang Zeng, Yihao Chen, Tianhe Ren, Junzhi Yu, and Lei Zhang. Detect anything via next point prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25472–25483, June 2026

    work page 2026

  78. [79]

    RaidaR: A Rich Annotated Image Dataset of Rainy Street Scenes

    Jiongchao Jin, Arezou Fatemi, Wallace Lira, Fenggen Yu, Biao Leng, Rui Ma, Ali Mahdavi-Amiri, and Hao Zhang. Raidar: A rich annotated image dataset of rainy street scenes, 2021. arXiv preprint arXiv:2104.04606

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  79. [80]

    Ultralytics datasets: Medical-pills detection dataset, Dec 2024

    Glenn Jocher and Muhammad Rizwan. Ultralytics datasets: Medical-pills detection dataset, Dec 2024. URL https: //docs.ultralytics.com/datasets/detect/medical-pills/

    work page 2024

  80. [81]

    Ultralytics datasets: Homeobjects-3k detection dataset, May 2025

    Glenn Jocher and Muhammad Rizwan. Ultralytics datasets: Homeobjects-3k detection dataset, May 2025. URL https: //docs.ultralytics.com/datasets/detect/homeobjects-3k/

    work page 2025

Showing first 80 references.