acceptodds
Withdrawn from ICLR 2027

When Visibility Degrades Together: Evaluating Human Pose Estimation across Severity and Compositional Low-Visibility Conditions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Human pose estimation (HPE) rarely operates under a single isolated visibility degradation, yet existing HPE evaluations typically treat each low-visibility condition as an independent fixed-difficulty bucket. They either isolate degradations into separate benchmarks or lack explicit low-visibility annotations, making severity-aware and compositional robustness analysis difficult. We introduce LV-Pose, a diagnostic evaluation framework for analyzing 2D HPE under severity-aware and compositional low-visibility conditions. LV-Pose reveals robustness failures that remain hidden under existing evaluation protocols. Individually mild degradations can compound when they co-occur, while performance within a single condition can deteriorate sharply as visibility severity increases. LV-Pose groups images into three difficulty levels (easy, medium, hard). Keypoint and severity annotations are provided for three primary conditions, haze/fog, low-light/nighttime, and precipitation, together with image-level tags for additional co-occurring visibility conditions. LV-Pose therefore supports severity-conditioned and composition-conditioned analysis of HPE robustness, instead of a single aggregate robustness claim under low visibility. Across representative HPE estimators and foundation models such as vision foundation models and vision-language models, LV-Pose reveals strong severity-dependent robustness degradation and compounding failures where co-occurring degradations reduce accuracy beyond their isolated effects. Restoration preprocessing also fails to consistently improve downstream pose robustness across severity levels, and can even degrade strong HPE systems despite visually enhanced inputs.

Primary area:
datasets and benchmarks
Keywords:
human pose estimation, low-visibility vision, robustness evaluation, benchmark, failure analysis, stress testing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_vSoqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposed a small but well curated dataset of 1,200 images that spans a variety of different challenging visual conditions for human pose estimation.

The paper is able to show that the ranking of state of the art models changes depending on the visual category and severity of the condition - this indicates existing benchmark that does not focus on the finegrained categorical performance of the pose estimator can not identify the best/most robust method.

Most interestingly to me, the paper examined the impact of image restoration methods as preprocessing on the pose estimator. They seem to have little effect if any, and depends a lot on the exact method being used and the exact pose estimator.

Critical strengths

The data collection process is rigorous in its design and well explained in the paper. The dataset diversity and quality is extensively evaluated and presented quite clearly.

The conclusions are well supported by the analysis and final experiments. The experiments are also extensive and well demonstrates the value of the dataset as a robustness benchmark.

Critical weaknesses

Since the methodology for creating the dataset is not novel and does not scale up easily, the paper's contribution is judged on the value of the dataset itself.

The dataset targets a small set of robustness relevant conditions: haze/fog, low-light/night, precipitation, glare/backlight, motion blur, focus blur. Since this selection of specific conditions is not well motivated, it is not clear to me how big of an impact this would be for the human pose estimation community.

Even if it is useful to understand the robustness of human pose estimation in these specific conditions, it is not clear how this analysis will improve and drive further algorithm development. The obvious thought is that we should scale up data collection/curation to target these less robust conditions. However this not very insightful to the learning community as a whole. The paper should demonstrate how these insights can be used to benefit model development beyond just data gathering to address these specific scenarios.

Questions for authors

Did I miss insights from the paper that goes beyond identifying the current data gap for training open source models?

The section on the image of restoration is interesting as an initial experimental validation that such methods don't make much difference, but it does not provide any insight on why this is the case. Further analysis on this question would have been more impactful to the image restoration field and better demonstrate the value of this dataset.

Reviewer_pET6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces LV-Pose, a benchmark designed to evaluate the robustness of 2D human pose estimation in low-visibility environments.

The authors have curated a dataset of 1,200 real-world images covering scenarios such as fog, low-light conditions, and precipitation. The dataset includes annotations for co-occurring degradations—such as glare and motion blur—as well as the severity levels of these various adverse conditions.

The study evaluates 14 pose estimation models and vision foundation models, and further examines the impact of various image restoration methods on pose estimation performance.

Critical strengths

There is only one critical strength: the author has meticulously curated a dataset of 1,200 real-world images captured in low-visibility conditions—a testament to the significant effort invested in establishing this benchmark. Other aspects, such as the clarity of the writing, have already been addressed in the designated selection fields.

Critical weaknesses
  1. My initial impression is that, for a high-caliber conference like ICLR, the volume of work in this paper is far too limited. It consists merely of collecting and annotating a dataset of 1,200 images. There is no algorithm, nor are there any novel scientific conclusions. It falls well short of the standards expected for ICLR.
  2. I agree with the premise of the paper; conditions such as fog, low light, rain, and snow indeed pose significant challenges for human pose estimation. However, it is highly regrettable that the authors provide only 2D annotations. In this era of world-model-driven research, I find it quite disappointing that a benchmark lacks any 3D annotations. The core difficulty in degraded environments—whether concerning the human body or the scene—lies in accurately recovering 3D information, which necessitates a deep understanding of depth and the relationship between foreground and background. This is where the true challenge of this topic resides, rather than merely annotating 2D human keypoints—especially using the basic COCO17 format. Moreover, I would particularly like to see 3D human annotations for the video.
  3. The authors do not propose a new pose estimation model or a new theory of robustness, nor do they provide an experimental design capable of accurately isolating the effects of various degradations. Furthermore, they fail to demonstrate how these annotations might improve existing 2D pose estimation methods.
Questions for authors

I also came across a paper titled "PoseBench" (2024)—which, incidentally, has not even been published yet. I am curious to know: compared to existing evaluation frameworks—such as PoseBench (2024) or LLIP (IJCAI 2025)—that already assess model robustness across various degradation types and severity levels, what unique insights does LV-Pose offer—beyond its annotations for multiple degradations in real-world images—that cannot be derived from these existing frameworks?

Reviewer_kHYgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposed Low-Visibility human pose estimiation (LV-Pose) for benchmarking human pose estimation under real world low-visibility conditions.

Critical strengths
  • The proposed benchmark covers a wide range of image degradation conditions, including haze/fog, low-light, and nighttime scenarios, revealing failure cases of existing haman pose estimation methods under these challenging conditions.
  • The annotation pipelie is easy to understand and well-designed, providing reasonable support for the quality and reliability of the annotations.
Critical weaknesses
  • The proposed benchmark contains only 411 images for haze/fog, 392 for low-light/night, and 397 for precipitation, which is relatively small compared to the scale of existing human pose estimation benchmarks. Given the limited scales of dataset in each degradation category, I am concerned that the benchmark may not adequately capture the diversity of challenging visiibility conditions, potentailly leading to over- or under-estimation of human pose estimation performance under such conditions.
  • Applying preprocessing techniques to enhance low-quality images is a common approach to improving downstream task performance. However, Table 3 shows that applying existing model-based image enhancement methods yield only limited performance improvements, while in some cases leading to substantial performance degradation. I suspect that this is becuase model-based enhancement methods generate artifacts, noise or distortions that negatively affect the performance of pose estimation models such as Sapiens2. Therefore, the authors should explan how and why image enhancement methods affects pose estimation performance.
  • The novelty of the paper is limited, as it focuses on evaluating and diagnosing the performance of exsting human pose estimation methods using the proposed benchmark. However, I think this work has the potential to make a more substantial contribution by scaling up the benchmark and investigating fine-tuning apporaches or training startegies that imporve the robustness of human pose estimation models under challenging visiblity conditions.
Questions for authors
  • Is the performance degradation of human pose estimation models under low-quality imaging conditions primarily due to the lack of training data, or does it stem from limitations in model architectures or training strategies that have not been adequately addressed?
  • Could the authors investigate how much performance improvement can be achieved on the proposed benchmark by training models such as Sapiens2 on existing human pose estimation datasets with conventional data augmentation, such as brightness reduction and other common image degradations?
  • Similarly, conducting comparative experiments with conventional image enhancement techniques, rather than model-based enhancement methods, could help quantify the extent to which HPE performance can be improved through simple preprocessing. Such experiments would also provide valuable insights into the limitations of existing approaches and help identify promising directions for benchmark expansion and model training strategies.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.