STAR-Pose: Efficient Low-Resolution Video Human Pose Estimation via Spatial-Temporal Adaptive Super-Resolution

Baojun Han; Furan An; Jinyan Chen; Yucheng Jin; Ziyue He

REVIEW 2 major objections 5 minor 30 references

STAR-Pose: Efficient Low-Resolution Video Human Pose Estimation via Spatial-Temporal Adaptive Super-Resolution

T0 review · 2 major / 5 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 A task-driven video super-resolution network beats generic upscaling-then-pose pipelines at extremely low resolution, gaining 5.2 mAP and roughly 3x speed.

desk verdict The reported efficiency numbers contradict the paper's own architecture description, so the speed advantage is unproven; the accuracy gains are plausible and the task-driven video SR idea deserves referee time. read the letter →

arxiv 2506.16061 v1 pith:XM7UOUOY submitted 2025-06-19 cs.CV

Yucheng Jin , Jinyan Chen , Ziyue He , Baojun Han , Furan An This is my paper

classification cs.CV

keywords humanposeestimationlow-resolutionvideosuper-resolutionlinearattentionspatial-temporaltransformertask-drivenlossefficientinference

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

STAR-Pose claims that human pose estimation from extremely low-resolution video is better done by a single end-to-end network that learns to super-resolve frames specifically for the pose task, rather than by upscaling first with a generic video super-resolver and then running pose estimation. On PoseTrack2018 at 64x48 input it reports 67.4% mAP, 5.2 points above the strongest cascade (DRCT+HRNet), while running at 32.4 ms per frame versus 93.1 ms. The network couples a spatial-temporal Transformer with linear attention to a lightweight CNN branch, and trains both with a pose-aware compound loss that mixes pixel reconstruction with keypoint heatmap supervision. If this transfers to real footage, task-driven super-resolution would be the cheaper and more accurate route for low-resolution video pose estimation.

What carries the argument

The mechanism is a dual-branch architecture. A spatial-temporal Transformer partitions video features into 3D patches, adds sinusoidal positional encodings along the temporal, width, and height axes, and replaces standard softmax attention with linear attention whose kernel is SCLeakyReLU — a shifted and clipped LeakyReLU with learnable truncation rate $\alpha$ and offset $\delta=1$ — bringing the complexity from $O(N^2)$ to $O(N)$ in sequence length. A parallel lightweight CNN branch built from Fused-MBConv and MBConv blocks extracts local texture, and an Adaptive Feature Fusion Module (AFFM) uses channel attention to combine the two branches at several stages before pixel-shuffle upsampling feeds a frozen HRNet-W48 pose head. The training objective is the pose-aware compound loss $L = \|I_{\mathrm{SR}} - I_{\mathrm{HR}}\|_1 + \lambda \|\hat{H} - H\|_2^2$ with $\lambda = 10$, which injects heatmap error into the super-resolution training so the network reconstructs structural features that help keypoint localization rather than maximizing PSNR.

What would settle it

Run the same comparison on genuine low-resolution video (surveillance or broadcast footage with native noise and compression) instead of synthetically downsampled crops and check whether the 5.2-point mAP lead over DRCT+HRNet at 64x48 survives; if it shrinks toward zero or reverses, the core transferability claim fails.

Watch

Extended reading notes

Core claim

STAR-Pose's central claim is that a task-oriented video super-resolution model, trained jointly with a pose estimator under a compound loss, outperforms both pose-only methods and cascades of generic video super-resolution plus pose estimation at very low resolutions, at a fraction of the inference cost. On PoseTrack2018, STAR-Pose reports 71.6% mAP at 128x96 and 67.4% mAP at 64x48, beating the strongest cascade (DRCT+HRNet) by 3.4 and 5.2 mAP points respectively, while running in 85.3 ms and 32.4 ms per frame versus 201.7 ms and 93.1 ms. The ablations back the design story: removing temporal modeling costs 4.5 mAP at 128x96 and 6.7 mAP at 64x48; replacing SCLeakyReLU with plain ReLU costs 1.8/1.1 mAP; removing the CNN branch and fusion costs 3.3-4.5 mAP; and removing the pose term from the loss costs 1.5-1.7 mAP even though PSNR rises by 0.26 dB. That last result is the paper's key evidence that pixel-level fidelity and keypoint localization are different objectives, and that the loss should target the pose task directly.

Load-bearing premise

The results assume that low-resolution video is faithfully simulated by Gaussian blurring and bicubic downsampling of sharp, well-cropped person videos; real footage with sensor noise, compression artifacts, motion blur, and imperfect detectors could erase the reported gains.

Editorial extensions

If this is right

At 64x48 inputs, a task-driven SR network can beat the best generic SR-plus-pose cascade by 5.2 mAP on PoseTrack2018, showing pixel fidelity is not the right proxy for pose accuracy.
The same architecture runs at 32.4 ms per frame versus 93.1 ms for DRCT+HRNet, roughly a 3x speedup, because super-resolution and pose estimation share one forward pass.
Temporal context is the largest single contributor: removing it drops mAP by 6.7 points at 64x48, so multi-frame information is essential for ultra-low-resolution pose.
Replacing the attention kernel with standard ReLU costs 1.1-1.8 mAP, meaning the negative-slope region of SCLeakyReLU carries useful signal in linear attention for this task.
Pose-supervision weight $\lambda$ must be balanced: raising it from 0.01 to 10 improves mAP while PSNR falls slightly, and beyond 10 both metrics drop.

Reading between the lines

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

Not stated in the paper, a likely extension: because the pose loss is computed against a frozen HRNet-W48 head, the optimized super-resolution features are probably tuned to that head's representation; swapping in a different pose network may require retraining the SR branch to keep the gain.
A testable extension the paper leaves implicit: the same task-driven compound-loss recipe could be applied to other dense prediction tasks at low resolution, such as semantic segmentation or face landmark detection, where pixel-only SR is known to be a poor proxy for downstream accuracy.
A practical implication the authors do not spell out: the efficiency gain comes from sharing one forward pass between SR and pose, so for deployment the advice is not 'better upscaling first' but 'a single network trained with a task loss.'
A caveat only implicit in the paper: the synthetic degradation pipeline omits compression artifacts and sensor noise, so the reported margins are upper-bound estimates; real-noise augmentation would be needed to preserve them in practice.

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.

Desk Editor's Note

read the letter

Quick take: the accuracy gains look plausible, but the efficiency claim doesn't survive arithmetic. In Table 1, STAR-Pose at 64x48 is 1.71 GFLOPs and 32.4 ms; HRNet-W48 alone at that resolution is 0.91 GFLOPs and 18.1 ms. That leaves ~0.8 GFLOPs and ~14 ms for the SR branch. But Section 3.1 says the fused features are pixel-shuffled upsampled before being fed to the HPE network, and the inputs are 4x downsampled, so HRNet should be seeing 256x192 and costing ~14.6 GFLOPs alone. The reported totals only hold if HRNet receives 64x48, which contradicts the described architecture. Either the GFLOPs/latency columns are miscalculated or the HRNet input is not what the text says. So the 2.8-4.4x speed advantage is not established from the paper as written.

What's worth credit: the idea of a task-driven video SR branch with a linear-attention transformer plus a CNN branch, and a pose-aware compound loss, is a reasonable, well-motivated combination for low-resolution video HPE. The ablation study is systematic: removing temporal context hurts the most, and the lambda sensitivity analysis shows a sensible accuracy/PSNR trade-off. The accuracy comparisons on PoseTrack 2018 and 2021 are consistent, with roughly 3-5% mAP gains over cascaded SR+HPE baselines at both resolutions. If those numbers hold, it's a useful engineering contribution.

Soft spots: all results are on validation sets, single-run, no error bars. The low-resolution degradation is synthetic (Gaussian blur plus bicubic downsampling) with ground-truth boxes, so real-world transfer is untested. No code or model released. The novelty is incremental — SCLeakyReLU is essentially LeakyReLU with a shift, and the architecture recombines EfficientViT, MBConv, and SRPose components — but for an applied paper that's acceptable.

Bottom line: the accuracy side deserves referee time; the speed side needs a correction and re-evaluation. I'd send it to reviewers with a specific request to verify the efficiency accounting, not desk-reject. I wouldn't cite it until the numbers are fixed.

Referee Report

2 major / 5 minor

Summary. The paper proposes STAR-Pose, a task-driven video super-resolution framework for low-resolution human pose estimation. The method combines a spatial-temporal Transformer with SCLeakyReLU linear attention, a parallel CNN branch, an adaptive feature fusion module, and a pose-aware compound loss. The HPE backbone is a frozen HRNet-W48. Experiments on PoseTrack2018 and PoseTrack2021 report mAP gains over SOTA and cascaded SR+HPE baselines at 128x96 and 64x48, along with claimed 2.8-4.4x faster inference. The main efficiency and accuracy claims depend on the consistency of Tables 1 and 3 and on the cascade evaluation protocol.

Significance. If the reported numbers are correct, the paper makes a useful contribution: it demonstrates that task-specific video super-resolution can improve low-resolution human pose estimation more efficiently than generic SR+HPE cascades. The component ablations in Table 4 and the lambda sensitivity analysis in Table 5 provide useful evidence for the design choices. However, the internal arithmetic inconsistency in the efficiency columns of Table 1 means that the speed advantage and the fairness of the cascade comparisons are not currently established. The core idea is interesting and worth pursuing, but the experimental accounting needs correction before the central claims can be accepted.

major comments (2)

[Section 3.1, Table 1, Table 3] The efficiency accounting is internally inconsistent. Section 3.1 and Figure 1 state that fused features are upsampled via pixel shuffle and then fed to the HPE network, yet the 64x48 row of Table 1 is only coherent if HRNet-W48 receives the 64x48 input: STAR-Pose totals 1.71 GFLOPs and 32.4ms, and subtracting the SR branch time of 14.3ms (Table 3) leaves 18.1ms, exactly the HRNet-at-64x48 baseline. If the pixel-shuffle upsampling is 4x, as implied by the 4x bicubic downsampling in Section 4.1, HRNet would process 256x192, costing about 14.6 GFLOPs and correspondingly more time; even a 2x upsampling would require 3.65 GFLOPs for HRNet alone, already exceeding the reported total. Please clarify the actual HRNet input resolution and recompute all GFLOPs/latency columns, or revise the architecture description. The claimed 2.8-4.4x speed advantage is not established from the current tables.
[Table 1, Table 3] The cascade baseline numbers at 64x48 appear to be sums of the SR time in Table 3 and the HRNet-at-64x48 time: DRCT+HRNet (93.1ms) equals DRCT (74.9ms) plus HRNet baseline (18.1ms), and SDBVSR+HRNet (142.5ms) equals SDBVSR (124.4ms) plus 18.1ms. This suggests that in these baselines the pose network received the low-resolution input rather than the super-resolved output, which would not be a genuine SR+HPE cascade. Please specify the exact protocol for each cascade (whether the SR model was pre-trained or fine-tuned, and at what resolution the HPE network received its input) and re-run the comparisons. This directly affects the validity of the reported accuracy advantage over cascaded approaches.

minor comments (5)

[Section 1 and Section 2.2] There are several typos and formatting inconsistencies: 'Howerver' in Section 1, 'representaion' in Section 1, 'a end-to-end' in Section 2.2, and inconsistent use of 'ST AR-Pose' in Table 1 and Table 2.
[Figure 3] The qualitative comparison in Figure 3 would be easier to interpret if the subfigures were labeled with the input resolution (128x96 or 64x48) and the pose keypoints were overlaid consistently across all panels.
[Section 4.3 and Table 4] All reported numbers are single-run results on validation sets; please state the experimental variance (e.g., multiple seeds or standard deviation) or explicitly note that no repeated runs were performed, especially since some ablation differences are close to 1% mAP.
[Section 4.1] The synthetic degradation protocol (Gaussian blur plus bicubic downsampling) and ground-truth bounding boxes are a simplification of real low-resolution video, which contains sensor noise, compression artifacts, motion blur, and detector errors. The paper should explicitly discuss this limitation and, if possible, include a test on a native low-resolution video benchmark or with detected boxes.
[Section 3.4 and Section 4.1] The phrase 'end-to-end' should be nuanced: the HRNet backbone is frozen during training, so the pose loss supervises only the SR branch. This is a valid task-driven training scheme, but it is not joint training of the pose estimator, and the text should say so explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim rests on direct pose-supervised training and independent benchmark comparisons, not on a self-referential derivation.

full rationale

STAR-Pose's derivation chain is self-contained. The core mechanism is end-to-end training with the compound loss L = ||I_SR - I_HR||_1 + lambda ||hat(H) - H||_2^2 (Eq. 7), where L_pose directly supervises predicted heatmaps against ground-truth heatmaps; this is standard supervised task-driven optimization, not a quantity defined in terms of the very result it is claimed to predict. The reported improvements (71.6 vs 68.2 mAP at 128x96 and 67.4 vs 62.2 at 64x48, Tables 1-2) are external benchmark numbers compared against published baselines, and the ablations in Table 4 remove components independently, so the contributions are not forced by construction. No load-bearing step reduces to a self-citation: references such as linear attention [13], HRNet [23], and SR backbones [11] are standard external building blocks or baselines, and no uniqueness theorem or author-defined ansatz is invoked to forbid alternatives. The lambda sensitivity analysis (Table 5) is validation-set selection rather than circularity, and the skeptic's GFLOPs/latency arithmetic issue in Table 1 is an internal consistency problem affecting the speed claim, not a case where a prediction equals its input by definition.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The central claim rests on standard linear-attention math, a synthetic degradation model, a fixed pretrained pose network used for supervision, and a single-person evaluation protocol. No new physical entities are introduced; SCLeakyReLU and AFFM are computational components whose value is only supported by in-paper ablations.

free parameters (3)

Compound loss weight lambda = 10
Chosen by validation-set search over {0.01, 0.1, 1, 5, 10, 25, 50, 100} (Table 5); the final reported mAP uses this tuned value on the same validation split.
SCLeakyReLU truncation rate alpha = learned, initialized 0.01
Truncation rate in the SCLeakyReLU activation; no theoretical derivation, learned during training.
Fusion gate beta = learned, initialized 0
Learnable parameter in Eq. 6 controlling how much CNN branch output is added to the transformer branch; no derivation.

assumptions (4)

standard math Associativity of matrix multiplication enables O(N*D^2) linear attention, reducing to O(N) when sequence length N >> feature dimension D.
Invoked in Section 3.2 to justify the efficiency claim; taken from Katharopoulos et al. [13].
domain assumption Gaussian blur plus bicubic downsampling is an adequate model of low-resolution video degradation.
Section 4.1 generates all low-res test inputs this way; real-world transfer is untested.
domain assumption A frozen HRNet-W48 pretrained on high-resolution images provides reliable heatmap supervision and final evaluation.
Section 3.4 and Section 4.1 use a fixed HRNet for both pose loss and evaluation; assumes its features are suitable for low-res SR guidance.
domain assumption Single-person crops at fixed resolution, derived from provided center and scale annotations, define the evaluation.
Section 4.1 uses top-down crops; the paper does not describe person detection at inference, so the protocol is a simplifying assumption.

invented entities (2)

SCLeakyReLU activation function
purpose: Non-negative activation with a small negative slope for linear attention, avoiding dying ReLU without exponentials
Only evidence is the in-paper ablation (b) in Table 4; no external benchmark or formal analysis.
Adaptive Feature Fusion Module (AFFM)
purpose: Channel-attention based fusion of transformer and CNN branch features
Validated only by ablation (c) in Table 4; no external analysis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of STAR-Pose: Efficient Low-Resolution Video Human Pose Estimation via Spatial-Temporal Adaptive Super-Resolution." pith.science (2026). https://pith.science/paper/XM7UOUOY

@misc{pith2026250616061,
  author       = {Pith},
  title        = {Pith review of: STAR-Pose: Efficient Low-Resolution Video Human Pose Estimation via Spatial-Temporal Adaptive Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XM7UOUOY}},
  note         = {Machine review of arXiv:2506.16061}
}

read the original abstract

Human pose estimation in low-resolution videos presents a fundamental challenge in computer vision. Conventional methods either assume high-quality inputs or employ computationally expensive cascaded processing, which limits their deployment in resource-constrained environments. We propose STAR-Pose, a spatial-temporal adaptive super-resolution framework specifically designed for video-based human pose estimation. Our method features a novel spatial-temporal Transformer with LeakyReLU-modified linear attention, which efficiently captures long-range temporal dependencies. Moreover, it is complemented by an adaptive fusion module that integrates parallel CNN branch for local texture enhancement. We also design a pose-aware compound loss to achieve task-oriented super-resolution. This loss guides the network to reconstruct structural features that are most beneficial for keypoint localization, rather than optimizing purely for visual quality. Extensive experiments on several mainstream video HPE datasets demonstrate that STAR-Pose outperforms existing approaches. It achieves up to 5.2% mAP improvement under extremely low-resolution (64x48) conditions while delivering 2.8x to 4.4x faster inference than cascaded approaches.

Figures

Figures reproduced from arXiv: 2506.16061 by the authors.

**Figure 1.** Overview of the STAR-Pose framework. It consists of three core components: (a) spatial-temporal transformer branch with linear attention; (b) parallel CNN branch; (c) adaptive feature fusion module (AFFM). 3.2 Efficient ST-Transformer with LeakyRelu Linear Attention Unlike traditional ViT[8] that divides 2D patches on single frames, we perform division in feature space and encode spatial-temporal positional informat… view at source ↗

**Figure 2.** Details of key components in STAR-Pose. (a) LeakyReLU-based linear spatialtemporal transformer architecture; (b) adaptive feature fusion module (AFFM). For computational efficiency, CNN branches employ Inverted Residual Blocks (MBConv) from MobileNetV3[10] and Fused-MBConv from EfficientNetV2[24]. Fused-MBConv merges expansion and depthwise convolutions into single standard convolutions at shallow layers. Specific… view at source ↗

**Figure 3.** Qualitative comparison of super-resolution outputs and corresponding pose estimation results across different methods. Implementation Details To systematically simulate low-resolution scenarios, we apply Gaussian blurring to original videos, followed by 2×/4× bicubic downsampling, generating two distinct resolution levels: low and ultra-low. Singleperson bounding boxes are adjusted to fixed resolutions of 128x96 a… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 27 canonical work pages

[1]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) Efficient Low-Resolution Video Human Pose Estimation 13

Andriluka, M., Iqbal, U., Insafutdinov, E., Pishchulin, L., Milan, A., Gall, J., Schiele, B.: PoseTrack: A benchmark for human pose estimation and tracking. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) Efficient Low-Resolution Video Human Pose Estimation 13

work page 2018
[2]

IEEE Transac- tions on Pattern Analysis and Machine Intelligence46(7), 4641–4653 (2024)

Bai, H., Pan, J.: Self-supervised deep blind video super-resolution. IEEE Transac- tions on Pattern Analysis and Machine Intelligence46(7), 4641–4653 (2024)

work page 2024
[3]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

Cai, H., Li, J., Hu, M., Gan, C., Han, S.: EfficientViT: Lightweight multi-scale attention for high-resolution dense prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17302–17313 (October 2023)

work page 2023
[4]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: BasicVSR++: Improving video super- resolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5972–5981 (June 2022)

work page 2022
[5]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22367–22377 (June 2023)

work page 2023
[6]

ACM Transactions on Embedded Computing Systems23(2) (Mar 2024)

Deng, J., Dong, S., Chen, L., Hu, J., Zhuo, C.: STDF: Spatio-temporal deformable fusion for video quality enhancement on embedded platforms. ACM Transactions on Embedded Computing Systems23(2) (Mar 2024)

work page 2024
[7]

In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

D¨ oring, A., Chen, D., Zhang, S., Schiele, B., Gall, J.: PoseTrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20963–20972 (June 2022)

work page 2022
[8]

arXiv preprint arXiv:2010.11929 (2020)

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

arXiv 2020

Show all 30 references

[9]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

He, J., Yang, W.: Video-based human pose regression via decoupled space-time aggregation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1022–1031 (June 2024)

2024
[10]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)

Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., Le, Q.V., Adam, H.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (October 2019)

2019
[11]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

Hsu, C.C., Lee, C.M., Chou, Y.S.: DRCT: Saving image super-resolution away from information bottleneck. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 6133–6142 (June 2024)

2024
[12]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)

Huang, J., Zhu, Z., Guo, F., Huang, G.: The devil is in the details: Delving into unbiased data processing for human pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2020)

2020
[13]

In: III, H.D., Singh, A

Katharopoulos, A., Vyas, A., Pappas, N., Fleuret, F.: Transformers are RNNs: Fast autoregressive transformers with linear attention. In: III, H.D., Singh, A. (eds.) Proceedings of the 37th International Conference on Machine Learning. Proceed- ings of Machine Learning Research...

2020
[14]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Kim, J., Oh, J., Lee, K.M.: Beyond image super-resolution for image recognition with task-driven perceptual loss. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2651–2661 (June 2024)

2024
[15]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Li, B., Li, X., Zhu, H., Jin, Y., Feng, R., Zhang, Z., Chen, Z.: SeD: Semantic- aware discriminator for image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 25784– 25795 (June 2024) 14 Jin et al

2024
[16]

IEEE Transactions on Image Processing 33, 2171–2182 (2024)

Liang, J., Cao, J., Fan, Y., Zhang, K., Ranjan, R., Li, Y., Timofte, R., Van Gool, L.: VRT: A video restoration transformer. IEEE Transactions on Image Processing 33, 2171–2182 (2024)

2024
[17]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops

Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: SwinIR: Image restoration using swin transformer. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. pp. 1833–1844 (October 2021)

2021
[18]

In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A

Liang, J., Fan, Y., Xiang, X., Ranjan, R., Ilg, E., Green, S., Cao, J., Zhang, K., Timofte, R., Gool, L.V.: Recurrent video restoration transformer with guided de- formable attention. In: Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., Oh, A. (eds.) Advances in Ne...

2022
[19]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Liu, Z., Chen, H., Feng, R., Wu, S., Ji, S., Yang, B., Wang, X.: Deep dual con- secutive network for human pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 525–534 (June 2021)

2021
[20]

arXiv preprint arXiv:1903.06733 (2019)

Lu, L., Shin, Y., Su, Y., Karniadakis, G.E.: Dying relu and initialization: Theory and numerical examples. arXiv preprint arXiv:1903.06733 (2019)

2019 arXiv
[21]

In: Jawahar, C.V., Li, H., Mori, G., Schindler, K

Neumann, L., Vedaldi, A.: Tiny people pose. In: Jawahar, C.V., Li, H., Mori, G., Schindler, K. (eds.) Computer Vision – ACCV 2018. pp. 558–574. Springer International Publishing, Cham (2019)

2019
[22]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2016)

Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Re...

2016
[23]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)

Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learn- ing for human pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)

2019
[24]

In: Meila, M., Zhang, T

Tan, M., Le, Q.: EfficientNetV2: Smaller models and faster training. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 10096–10106. PMLR (18–24 Jul 2021)

2021
[25]

ACM Comput

Tay, Y., Dehghani, M., Bahri, D., Metzler, D.: Efficient transformers: A survey. ACM Comput. Surv.55(6) (Dec 2022)

2022
[26]

In: Proceedings of the 31st ACM International Conference on Multimedia

Wang, H., Liu, J., Tang, J., Wu, G.: Lightweight super-resolution head for human pose estimation. In: Proceedings of the 31st ACM International Conference on Multimedia. p. 2353–2361. MM ’23, Association for Computing Machinery, New York, NY, USA (2023)

2023
[27]

Pattern Recognition150, 110352 (2024)

Wang, S., Sang, Y., Liu, Y., Wang, C., Lu, M., Sun, J.: Prior based pyramid residual clique network for human body image super-resolution. Pattern Recognition150, 110352 (2024)

2024
[28]

In: Peng, Y., Hu, S.M., Gabbouj, M., Zhou, K., Elad, M., Xu, K

Xu, J., Liu, Y., Zhao, L., Zhang, S., Yang, J.: Tiny person pose estimation via image and feature super resolution. In: Peng, Y., Hu, S.M., Gabbouj, M., Zhou, K., Elad, M., Xu, K. (eds.) Image and Graphics. pp. 315–327. Springer International Publishing, Cham (2021)

2021
[29]

Computers & Electrical Engineering 93, 107192 (2021)

Zhang, Z., Wan, L., Xu, W., Wang, S.: Estimating a 2d pose from a tiny person image with super-resolution reconstruction. Computers & Electrical Engineering 93, 107192 (2021)

2021
[30]

Machine Learning114(6), 135 (2025)

Zhang, Z., Wan, L., Xu, W., Wang, S.: Low-resolution human pose estimation and action recognition via pose-driven super-resolution reconstruction. Machine Learning114(6), 135 (2025)

2025

Pith tools

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

[1] [1]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) Efficient Low-Resolution Video Human Pose Estimation 13

Andriluka, M., Iqbal, U., Insafutdinov, E., Pishchulin, L., Milan, A., Gall, J., Schiele, B.: PoseTrack: A benchmark for human pose estimation and tracking. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2018) Efficient Low-Resolution Video Human Pose Estimation 13

work page 2018

[2] [2]

IEEE Transac- tions on Pattern Analysis and Machine Intelligence46(7), 4641–4653 (2024)

Bai, H., Pan, J.: Self-supervised deep blind video super-resolution. IEEE Transac- tions on Pattern Analysis and Machine Intelligence46(7), 4641–4653 (2024)

work page 2024

[3] [3]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

Cai, H., Li, J., Hu, M., Gan, C., Han, S.: EfficientViT: Lightweight multi-scale attention for high-resolution dense prediction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 17302–17313 (October 2023)

work page 2023

[4] [4]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chan, K.C., Zhou, S., Xu, X., Loy, C.C.: BasicVSR++: Improving video super- resolution with enhanced propagation and alignment. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5972–5981 (June 2022)

work page 2022

[5] [5]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22367–22377 (June 2023)

work page 2023

[6] [6]

ACM Transactions on Embedded Computing Systems23(2) (Mar 2024)

Deng, J., Dong, S., Chen, L., Hu, J., Zhuo, C.: STDF: Spatio-temporal deformable fusion for video quality enhancement on embedded platforms. ACM Transactions on Embedded Computing Systems23(2) (Mar 2024)

work page 2024

[7] [7]

In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

D¨ oring, A., Chen, D., Zhang, S., Schiele, B., Gall, J.: PoseTrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20963–20972 (June 2022)

work page 2022

[8] [8]

arXiv preprint arXiv:2010.11929 (2020)

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

arXiv 2020