Pith. sign in

REVIEW 4 major objections 6 minor 29 references

ParkFormer: A Transformer-Based Parking Policy with Goal Embedding and Pedestrian-Aware Control

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An end-to-end Transformer parking policy reaches 96.57% success in CARLA by fusing goal-slot attention with predicted pedestrian motion.

desk verdict A useful incremental parking policy with solid ablations, but the headline comparison to E2E Parking is undermined by an input mismatch: ParkFormer gets goal and pedestrian states as oracle inputs, so the claimed gains may be an information advantage, not the architecture. read the letter →

arxiv 2506.16856 v1 pith:AK5DDLMS submitted 2025-06-20 cs.CV cs.AI

Jun Fu , Bin Tian , Haonan Chen , Shi Meng , Tingting Yao This is my paper
classification cs.CVcs.AI
keywords autonomousparkingend-to-enddrivingTransformerbird's-eyeviewpedestrianpredictioncross-attentionCARLAsimulatorimitationlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a single Transformer-based end-to-end policy can learn to park from expert demonstrations in both perpendicular and parallel slots while reacting to pedestrians, without the modular perception-planning-control pipeline of conventional systems. In closed-loop CARLA experiments, it reports a 96.57% success rate with 0.21 m position error and 0.41 degree orientation error in vertical parking, and 97.30% success in parallel parking, beating the E2E Parking baseline on every reported metric. The authors attribute the result to a goal-conditioned cross-attention mechanism that focuses bird's-eye-view features on the target slot and to a GRU-based pedestrian predictor that lets the policy anticipate and yield to moving people. If the claims hold, imitation-learned driving policies become a viable alternative to hand-designed parking stacks in constrained, dynamic environments.

What carries the argument

The central mechanism is a goal-conditioned cross-attention module: the target slot embedding $G=(x_g,y_g,\psi_g)$ is projected through an MLP and used as the query, while the bird's-eye-view feature map supplies keys and values, so the model selectively emphasizes visual regions relevant to reaching the specified slot. Around it sit a GRU-based pedestrian predictor that forecasts ten steps of pedestrian motion and feeds an ego-pedestrian cross-attention layer, plus an autoregressive Transformer decoder that emits discrete control tokens as a sequence. Together these turn heterogeneous slot geometries and dynamic obstacles into one goal-aware, motion-aware representation from which smooth control sequences are generated.

What would settle it

Run the trained policy in closed loop while adding Gaussian noise with standard deviation around 0.3 m to the target slot center and randomly clearing the pedestrian mask for a fraction of frames; if success drops markedly or collision rate rises well above 1.16%, the perfect-information assumption is the load-bearing part of the result.

Watch

Extended reading notes

Core claim

The central claim is that end-to-end imitation learning, rather than explicit modeling, can produce a parking policy that is both precise and socially aware. The network consumes surround-view RGB images, the target slot state $G=(x_g,y_g,\psi_g)$, ego velocity and acceleration, and pedestrian positions, velocities, accelerations, and existence masks, and outputs discrete control tokens for throttle, braking, steering, and gear through an autoregressive Transformer decoder. In the CARLA Town04-Opt vertical scenario it reports 96.57% success, 0.21 m average position error, 0.41 degree average orientation error, and a 1.16% collision rate; in Town10HD-Opt parallel parking it reports 97.30% success with 0.25 m and 0.43 degree errors and 1.40% collision rate. The ablation study shows that removing the pedestrian predictor raises collision rate to 20.54%, and replacing target cross-attention with simple concatenation raises position error to 0.92 m, which supports the paper's attribution of safety to pedestrian modeling and precision to goal-conditioned attention.

Load-bearing premise

The load-bearing premise is that the target slot state $G=(x_g,y_g,\psi_g)$ and the pedestrian state $P$ are given to the policy exactly as provided, so the reported success rate measures the controller under perfect information rather than a full camera-to-control perception pipeline.

Editorial extensions

If this is right

  • A single model can handle both perpendicular and parallel slot geometries once the goal is embedded as a shared low-dimensional query signal, removing the need for separate policies per slot type.
  • Explicit pedestrian trajectory prediction has a large measured effect on safety: collision rate jumps from 1.16% to 20.54% when the pedestrian module is removed.
  • Replacing goal cross-attention with naive concatenation degrades final positioning, indicating that the way the goal is fused matters at least as much as having the goal information at all.
  • Discretizing control into tokens and predicting them autoregressively with a Transformer produces temporally coherent maneuvers, since each new control is conditioned on the previously emitted ones.
  • The released CARLA dataset with randomized pedestrians gives the community a common benchmark for comparing future end-to-end parking policies.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the target slot state and pedestrian states are fed into the network as clean values, the 96.57% success rate is a policy upper bound; a real system must first detect both from images, and noisy detections will likely erode this number.
  • Editorial inference: the goal-embedding-by-query design could transfer to other maneuvering tasks such as garage entry, docking, or diagonal and oblique parking by changing the goal vector, without retraining the perception stack.
  • Editorial inference: the GRU pedestrian predictor is trained with teacher forcing, so its errors may accumulate over the ten-step horizon at inference; evaluating the policy with predicted rather than ground-truth pedestrian trajectories would test whether the safety gain persists.
  • Editorial inference: because the simulator scenes remove default CARLA objects and use controlled pedestrian spawns, the model's response to clutter, weather, and lighting variation in real deployment remains untested by the reported experiments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes ParkFormer, a Transformer-based end-to-end parking policy trained by imitation learning on CARLA demonstrations. The network takes surround-view camera images, a target-slot pose G, ego-vehicle motion, and pedestrian dynamics as inputs, and outputs an autoregressive sequence of discretized control tokens. Its main design elements are a goal-conditioned cross-attention module that fuses BEV features with the target slot embedding and a GRU-based pedestrian predictor. The authors evaluate closed-loop parking performance in two CARLA scenarios (Town04-Opt vertical and Town10HD-Opt parallel) and report a success rate of 96.57%, position error of 0.21 m, and orientation error of 0.41 degrees. They also compare against E2E Parking [25] and present ablations for the pedestrian prediction module and the target fusion mechanism.

Significance. If the comparative claim is substantiated, the proposed architecture is a credible end-to-end parking policy that combines goal-conditioned spatial attention with pedestrian-aware control in a single trainable model, and the planned release of code and a dynamic parking dataset would be a useful resource for the community. The closed-loop evaluation with a pedestrian-aware module addresses a genuinely relevant safety aspect of autonomous parking. However, the central quantitative claims are currently not fully verified: the comparison with E2E Parking does not establish input equivalence, the evaluation lacks trial counts and statistical confidence, and the parallel-scenario result has no baseline. The paper is therefore a useful incremental contribution whose headline numbers still need supporting evidence.

major comments (4)
  1. [Section IV-D, Table I] Input equivalence between ParkFormer and E2E Parking is not established. Equation (1) gives ParkFormer privileged access to the goal pose G and pedestrian dynamics P, while Section II-B describes E2E Parking as mapping 'images and motion states' to control signals. Section IV-D states only that dataset split, perception backbone, and optimizer hyperparameters are identical; it does not state whether E2E Parking was also given G and/or P. The claimed 30% and 58.6% reductions in PE and OE in Table I are therefore not attributable to the proposed architecture unless an input-equivalent baseline is provided. Please add an E2E Parking variant that receives G and P, or explicitly add an ablation that removes G from ParkFormer, and report the comparison under identical input conditions.
  2. [Section IV-D, Table I] The evaluation reports no number of closed-loop trials, no standard deviations, and no statistical significance tests. Given that the entire dataset consists of 272 episodes (Section IV-A), the reported margin between 91.41% and 96.57% success rates could plausibly lie within run-to-run or episode-to-episode variation. Please report the number of test episodes per scenario and per condition, along with confidence intervals or repeated-seed results.
  3. [Section IV-D, Table I] No baseline result is reported for Town10HD-Opt: the E2E Parking row is marked with '-'. The text nevertheless claims that the parallel-scenario results demonstrate robustness and generalization. Without a baseline in the parallel scenario, the comparative claim is restricted to Town04-Opt; please either run the baseline in Town10HD-Opt or temper the generalization claim.
  4. [Section IV-A] The dataset retains only trajectories whose final position and yaw errors are below 0.5 m and 0.5 degrees. This quality filter may remove the hardest demonstrations and make the learning problem easier; the paper does not report how many trajectories were discarded. Please quantify the filtering rate and discuss its effect on the reported success rate and error metrics.
minor comments (6)
  1. [Abstract / Section IV-D] The text states that OE is reduced by approximately 58.6%, but Table I reports 0.87 deg for E2E Parking and 0.41 deg for ParkFormer, which is a reduction of 52.9%; please correct the percentage or the table.
  2. [Sections IV-A and IV-B] The dataset size is reported as approximately 46,400 frames at 10 Hz in Section IV-A and approximately 45,000 multimodal frames in Section IV-B; please make the two counts consistent.
  3. [Section III-D] The pedestrian prediction horizon of 10 steps is stated without the temporal step size or the loss function used to train the predictor; specifying both is necessary to reproduce this module.
  4. [Section III-A, Eq. (1)] P_{i,j} is defined as a tuple of positions, velocities, accelerations, and a binary mask, but the text also refers to multiple pedestrians; please clarify how multiple pedestrians are packed into a fixed-size input and how the mask is used.
  5. [References] References [22] and [27] appear to be the same RL-OGM-Parking entry listed twice; please remove the duplicate and fix the in-text citation numbering.
  6. [References] Reference [30] (Pix2seq) does not appear to be cited in the body of the paper; please either cite it or remove it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported success rates and errors are empirical closed-loop measurements, not quantities forced by the paper's training objective or by self-citation.

full rationale

ParkFormer's central claims are empirical results from CARLA closed-loop experiments. The training loss in Eq. 4 is a cross-entropy loss on discretized expert control tokens, and none of the reported metrics (SR, PE, OE, CR) appear as terms in that loss; therefore the reported numbers are not fitted parameters renamed as predictions. The dataset is collected from expert demonstrations with quality filtering, but the model is evaluated on held-out episodes, so the success rate is a genuine measurement rather than a construction. The comparison with E2E Parking raises a possible experimental fairness concern, because ParkFormer receives privileged inputs G and P that E2E Parking may not receive, but that is an issue of input equivalence and benchmark design, not a circular derivation: the paper's own equations do not reduce its conclusion to its inputs. The paper does not rely on load-bearing self-citations, does not invoke any uniqueness theorem from the authors' prior work, and does not rename a known result. The ablation studies are independent empirical tests of module contributions. I find no step in the paper's derivation chain that is circular by the criteria of this analysis, so the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The central claim rests on standard deep learning components, three hand-chosen hyperparameters (binning boundaries, prediction horizon, BEV grid), and three domain assumptions: CARLA realism, exact oracle state inputs, and sufficiency of LSS BEV.

free parameters (3)
  • Control token bin boundaries = not reported
    The paper discretizes acceleration, steering, and gear into 'finite token spaces through adaptive binning strategies optimized for vehicle dynamics' (Section III-E). The bin widths are chosen by hand, and the exact values are not provided, so the precision and smoothness of control depend on an unreported choice.
  • Pedestrian prediction horizon = 10 steps
    The GRU predictor forecasts 10 future steps (Section III-D). The horizon is chosen by hand, not justified by data, and directly affects how early the vehicle can react.
  • BEV grid resolution = 200x200 at 0.1 m
    The BEV representation covers 20 m x 20 m at 0.1 m resolution (Section IV-B). This choice controls spatial precision; a coarser grid could reduce placement accuracy.
assumptions (3)
  • domain assumption CARLA 0.9.14 is a valid proxy for real parking environments
    All results are obtained in CARLA with default objects removed and manually designed layouts (Section IV-A). The paper states generalization to real world remains to be explored (Section V).
  • domain assumption Exact target slot state and pedestrian states are available at inference
    The problem formulation (Section III-A, Eq. 1) defines G and P as direct inputs. The network never learns to detect slots or pedestrians from images, so the high success rate presumes perfect perception of these states.
  • standard math LSS depth estimation from surround cameras produces a sufficient BEV representation
    The BEV features are built with the LSS method [29] (Section III-C). The paper does not quantify error in the BEV reconstruction or its effect on parking.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ParkFormer: A Transformer-Based Parking Policy with Goal Embedding and Pedestrian-Aware Control." pith.science (2026). https://pith.science/paper/AK5DDLMS

@misc{pith2026250616856,
  author       = {Pith},
  title        = {Pith review of: ParkFormer: A Transformer-Based Parking Policy with Goal Embedding and Pedestrian-Aware Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AK5DDLMS}},
  note         = {Machine review of arXiv:2506.16856}
}
read the original abstract

Autonomous parking plays a vital role in intelligent vehicle systems, particularly in constrained urban environments where high-precision control is required. While traditional rule-based parking systems struggle with environmental uncertainties and lack adaptability in crowded or dynamic scenes, human drivers demonstrate the ability to park intuitively without explicit modeling. Inspired by this observation, we propose a Transformer-based end-to-end framework for autonomous parking that learns from expert demonstrations. The network takes as input surround-view camera images, goal-point representations, ego vehicle motion, and pedestrian trajectories. It outputs discrete control sequences including throttle, braking, steering, and gear selection. A novel cross-attention module integrates BEV features with target points, and a GRU-based pedestrian predictor enhances safety by modeling dynamic obstacles. We validate our method on the CARLA 0.9.14 simulator in both vertical and parallel parking scenarios. Experiments show our model achieves a high success rate of 96.57\%, with average positional and orientation errors of 0.21 meters and 0.41 degrees, respectively. The ablation studies further demonstrate the effectiveness of key modules such as pedestrian prediction and goal-point attention fusion. The code and dataset will be released at: https://github.com/little-snail-f/ParkFormer.

Figures

Figures reproduced from arXiv: 2506.16856 by the authors.

Figure 1
Figure 1. Overview of the proposed framework. The sys [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of the proposed Transformer-based end-to-end framework for autonomous parking in dynamic [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. (a)–(c) and (e)–(g) illustrate BEV views from [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 27 canonical work pages

  1. [25]

    Y . Yang, D. Chen, T. Qin, X. Mu, C. Xu and M. Yang, ”E2E Parking: Autonomous Parking by the End-to-end Neural Network on the CARLA Simulator,” 2024 IEEE Intelligent Vehicles Symposium (IV), Jeju Island, Korea, Republic of, 2024, pp. 2375-2382

    work page 2024

  2. [1]

    Y . Wang, E. Hansen and H. Ahn, ”Hierarchical Planning for Au- tonomous Parking in Dynamic Environments,” in IEEE Transactions on Control Systems Technology, vol. 32, no. 4, pp. 1386-1398, July 2024

    work page 2024

  3. [2]

    Codevilla, M

    F. Codevilla, M. M ¨uller, A. L ´opez, V . Koltun and A. Dosovitskiy, ”End-to-End Driving Via Conditional Imitation Learning,” 2018 IEEE International Conference on Robotics and Automation (ICRA), Bris- bane, QLD, Australia, 2018, pp. 4693-4700

    work page 2018

  4. [3]

    Trajectoryguided control prediction for end-to-end autonomous driving: A simple yet strong baseline,

    P. Wu, X. Jia, L. Chen, J. Yan, H. Li, and Y . Qiao, “Trajectoryguided control prediction for end-to-end autonomous driving: A simple yet strong baseline,” Advances in Neural Information Processing Systems, vol. 35, pp. 6119–6132, 2022

    work page 2022

  5. [4]

    X. Mu, T. Qin, S. Zhang, C. Xu and M. Yang, ”Pix2Planning: End- to-End Planning by Vision-language Model for Autonomous Driving on Carla Simulator,” 2024 IEEE Intelligent Vehicles Symposium (IV), Jeju Island, Korea, Republic of, 2024, pp. 2383-2390

    work page 2024

  6. [5]

    Attention is all you need,

    A. Vaswani et al., “Attention is all you need,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

    work page 2017

  7. [6]

    Nayakanti, R

    N. Nayakanti, R. Al-Rfou, A. Zhou, K. Goel, K. S. Refaat and B. Sapp, ”Wayformer: Motion Forecasting via Simple & Efficient Attention Networks,” 2023 IEEE International Conference on Robotics and Automation (ICRA), London, United Kingdom, 2023, pp. 2980- 2987

    work page 2023

  8. [7]

    Z. Li et al., ”BEVFormer: Learning Bird’s-Eye-View Representation From LiDAR-Camera via Spatiotemporal Transformers,” in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 3, pp. 2020-2036, March 2025

    work page 2020

Show all 29 references
  1. [8]

    Z. Liu et al., ”BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird’s-Eye View Representation,” 2023 IEEE International Conference on Robotics and Automation (ICRA), London, United Kingdom, 2023, pp. 2774-2781

    2023

  2. [9]

    ”Vad: Vectorized scene representation for efficient autonomous driving.” Proceedings of the IEEE/CVF International Conference on Computer Vision

    Jiang, Bo, et al. ”Vad: Vectorized scene representation for efficient autonomous driving.” Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023

    2023

  3. [10]

    ”Vadv2: End-to-end vectorized autonomous driving via probabilistic planning.” arXiv preprint arXiv:2402.13243 (2024)

    Chen, Shaoyu, et al. ”Vadv2: End-to-end vectorized autonomous driving via probabilistic planning.” arXiv preprint arXiv:2402.13243 (2024)

    2024 arXiv

  4. [11]

    ”FSGA: Motion Prediction Method Based on Future Scene Graph Attention.” IEEE Transactions on Intelligent Transporta- tion Systems (2025)

    Xu, Nan, et al. ”FSGA: Motion Prediction Method Based on Future Scene Graph Attention.” IEEE Transactions on Intelligent Transporta- tion Systems (2025)

    2025

  5. [12]

    ”Hivt: Hierarchical vector transformer for multi- agent motion prediction.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhou, Zikang, et al. ”Hivt: Hierarchical vector transformer for multi- agent motion prediction.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022

    2022

  6. [13]

    ”Rctrans: Radar-camera trans- former via radar densifier and sequential decoder for 3d object detec- tion.” Proceedings of the AAAI Conference on Artificial Intelligence

    Li, Yiheng, Yang Yang, and Zhen Lei. ”Rctrans: Radar-camera trans- former via radar densifier and sequential decoder for 3d object detec- tion.” Proceedings of the AAAI Conference on Artificial Intelligence. V ol. 39. No. 5. 2025

    2025

  7. [14]

    Shamsoshoara, S

    A. Shamsoshoara, S. B. Salih and P. Aghazadeh, ”SwapTransformer: Highway Overtaking Tactical Planner Model via Imitation Learning on OSHA Dataset,” in IEEE Access, vol. 12, pp. 76194-76206, 2024

    2024

  8. [15]

    Z. Chen, Z. Yu, J. Li, L. You and X. Tan, ”DualAT: Dual Attention Transformer for End-to-End Autonomous Driving,” 2024 IEEE Inter- national Conference on Robotics and Automation (ICRA), Yokohama, Japan, 2024, pp. 16353-16359

    2024

  9. [16]

    Azam and V

    S. Azam and V . Kyrki, ”Multi-Task Adaptive Gating Network for Trajectory Distilled Control Prediction,” in IEEE Robotics and Au- tomation Letters, vol. 9, no. 5, pp. 4862-4869, May 2024

    2024

  10. [17]

    ”Planning-oriented autonomous driving.” Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion

    Hu, Yihan, et al. ”Planning-oriented autonomous driving.” Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion. 2023

    2023

  11. [18]

    ”Genad: Generative end-to-end autonomous driving.” European Conference on Computer Vision

    Zheng, Wenzhao, et al. ”Genad: Generative end-to-end autonomous driving.” European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024

    2024

  12. [19]

    ”Think twice before driving: Towards scalable decoders for end-to-end autonomous driving.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Jia, Xiaosong, et al. ”Think twice before driving: Towards scalable decoders for end-to-end autonomous driving.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023

    2023

  13. [20]

    ”A reinforce- ment learning-based reverse-parking system for autonomous vehicles.” IET Intelligent Transport Systems 19.1 (2025): e12614

    Al-Mousa, Amjed, Ahmad Arrabi, and Hamza Daoud. ”A reinforce- ment learning-based reverse-parking system for autonomous vehicles.” IET Intelligent Transport Systems 19.1 (2025): e12614

    2025

  14. [21]

    Selvaraj, R

    S. Selvaraj, R. Thangarajan, J. Regiin Arul Rithik, S. Pranesan and M. Karthika, ”Intelligent Parking Assistance Using Deep Rein- forcement Learning and YOLOv8 in Simulated Environments,” 2025 International Conference on Multi-Agent Systems for Collaborative Intelligence (ICMS...

    2025

  15. [23]

    X. Shen, I. Batkovic, V . Govindarajan, P. Falcone, T. Darrell and F. Borrelli, ”ParkPredict: Motion and Intent Prediction of Vehicles in Parking Lots,” 2020 IEEE Intelligent Vehicles Symposium (IV), Las Vegas, NV , USA, 2020, pp. 1170-1175

    2020

  16. [24]

    X. Shen, M. Lacayo, N. Guggilla and F. Borrelli, ”ParkPredict+: Multimodal Intent and Motion Prediction for Vehicles in Parking Lots with CNN and Transformer,” 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC), Macau, China, 2022, pp. 3999-4004

    2022

  17. [26]

    C. Li, Z. Ji, Z. Chen, T. Qin and M. Yang, ”ParkingE2E: Camera- based End-to-end Parking Network, from Images to Planning,” 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Abu Dhabi, United Arab Emirates, 2024, pp. 13206-13212

    2024

  18. [27]

    ”RL-OGM-Parking: Lidar OGM-based hybrid re- inforcement learning planner for autonomous parking.” arXiv preprint arXiv:2502.18846 (2025)

    Wang, Zhitao, et al. ”RL-OGM-Parking: Lidar OGM-based hybrid re- inforcement learning planner for autonomous parking.” arXiv preprint arXiv:2502.18846 (2025)

    2025 arXiv

  19. [28]

    ”TransParking: A Dual-Decoder Transformer Framework with Soft Localization for End-to-End Auto- matic Parking.” arXiv preprint arXiv:2503.06071 (2025)

    Du, Hangyu, and Chee-Meng Chew. ”TransParking: A Dual-Decoder Transformer Framework with Soft Localization for End-to-End Auto- matic Parking.” arXiv preprint arXiv:2503.06071 (2025)

    2025

  20. [29]

    Philion, Jonah, and Sanja Fidler. ”Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d.” Com- puter Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16. Springer International Publish...

    2020

  21. [30]

    ”Pix2seq: A language modeling framework for object detection.” arXiv preprint arXiv:2109.10852 (2021)

    Chen, Ting, et al. ”Pix2seq: A language modeling framework for object detection.” arXiv preprint arXiv:2109.10852 (2021)

    2021 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.