Occluded Gait Recognition with Mixture of Experts: An Action Detection Perspective
Panjian Huang, Yunjie Peng, Saihui Hou, Chunshui Cao, Xu Liu, Zhiqiang He, Yongzhen Huang
cs.CV
Quick verdict
GaitMoE recasts gait as action composition via temporal/action experts and introduces OccGait; strong results, but thin margins lack significance testing.
Summary
The paper proposes GaitMoE, an occluded gait recognition model that recasts a gait sequence as a composition of actions, borrowing from temporal action detection. A Mixture of Temporal Experts (MTE) applies dilated convolutions with different dilation ratios to channel segments to build action anchors, and a Mixture of Action Experts (MAE) uses attention-like scoring to integrate similar anchors into action proposals; the model is trained end-to-end with only identity labels (triplet + cross-entropy). The authors also introduce OccGait, a real-captured benchmark with 101 subjects, 8 views, over 80k sequences, and four annotated occlusion types (none, carrying, crowd, static). GaitMoE reports Rank-1 of 91.4 (NM) and 82.1 (CA) on OccGait, 85.3 mean on OccCASIA-B, and state-of-the-art 73.7 Rank-1 / 66.2 mAP on Gait3D and 79.6 Rank-1 on GREW (GaitMoE-B), at moderate cost (27.75M params, 36.95G FLOPs for GaitMoE-T). Ablations (Tabs. 4-5) support both modules and the choice of 16 action experts. Main limitations: on OccGait crowd occlusion, GaitMoE loses to GaitBase at several views (e.g., 57.6 vs 60.0 at 180 degrees), all results are single runs without significance tests, baseline comparisons on the occluded benchmarks omit stronger recent methods (DyGait, HSTL), and only the dataset, not method code, is explicitly stated as released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Dynamic contextual information within gait sequences (gait continuity across adjacent frames and gait cycles) can infer and integrate occluded body regions.
- Casting gait recognition as temporal action detection, with Mixture of Temporal Experts building action anchors and Mixture of Action Experts integrating them into action proposals, trains end-to-end with only ID labels and improves occlusion robustness.
- OccGait is the first publicly released gait benchmark combining diverse real occlusion scenarios with explicit occlusion-type annotations.
- GaitMoE achieves state-of-the-art accuracy on OccGait, OccCASIA-B, Gait3D and GREW at moderate computational cost.
Strengths
New real-scenario occlusion benchmark with explicit annotations
OccGait (Sec. 4) offers 101 subjects, 8 views, >80k sequences, four annotated conditions (NM/CA/CR/ST) with defined probe/gallery protocol; informed consent and privacy are addressed — filling a genuine gap vs. synthetic occlusion datasets.
Coherent motivation-to-design mapping
Sec. 1 links gait continuity and gait cycles to occlusion robustness, and Secs. 3.1-3.2 instantiate these as action anchors (dilated temporal experts over channel segments, Eq. 1) and action proposals (softmax-scored weighted integration, Eqs. 2-3), trained end-to-end with ID labels only.
Consistent gains across four benchmarks
Favorable accuracy-efficiency trade-off
Modular ablations and interpretability
Weaknesses
No statistical rigor
Crowd-occlusion claim contradicted per-view
Possible preprocessing asymmetry
Baseline coverage gaps on occluded benchmarks
MoE framing partly rhetorical
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is GaitMoE and how does it handle occlusion in gait recognition?
GaitMoE treats a gait video as a composition of actions instead of raw frames. Its Mixture of Temporal Experts uses dilated convolutions at multiple temporal scales to build action anchors, and its Mixture of Action Experts scores and merges similar anchors (spanning different gait cycles) into robust action proposals. Because whole gait cycles repeat, information from unoccluded cycles can compensate for occluded ones. It is trained end-to-end using only person-identity labels.
What is the OccGait dataset?
OccGait is a real-captured occluded gait benchmark with 101 subjects, 8 camera views, and over 80,000 sequences. Each subject is recorded under four annotated occlusion types: no occlusion, carrying occlusion (umbrella, luggage), crowd occlusion (walking with another person), and static occlusion (plants, chairs). The standard protocol uses NM01/NM02 sequences as the gallery and reports Rank-1 accuracy; the authors state it is available at github.com/BNU-IVC/OccGait.
How accurate is GaitMoE compared to other gait recognition methods?
The paper reports Rank-1 accuracy of 91.4% (normal walking) and 82.1% (carrying) on OccGait, a mean of 85.3% on OccCASIA-B, 73.7% Rank-1 / 66.2% mAP on Gait3D, and 79.6% Rank-1 on GREW (GaitMoE-B variant), which exceeds prior methods like DyGait (66.3% on Gait3D) at lower computational cost. Note these are single-run results without significance testing, and on crowd occlusion some per-view results are below GaitBase.
Is the GaitMoE code or the OccGait dataset publicly available?
The paper explicitly states that the OccGait database is available at https://github.com/BNU-IVC/OccGait. A public code release for the GaitMoE model itself is not explicitly stated in the provided text, so reproducibility currently rests on the detailed hyperparameters in Sec. 5.2.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-7 are available only as text descriptions; qualitative claims about visualizations and the efficiency plot could not be independently verified.
- Supplementary material (e.g., full S,K ablation details, architecture specifics) is referenced but not included in the extracted text.
- Only the OccGait dataset release is explicitly stated; method code availability is inferred as not stated.
- This is an arXiv preprint (2609.18432, LNCS-style formatting); peer-review status is unknown.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.