Gait Recognition in the Wild: A Large-Scale Benchmark and NAS-Based Baseline
Summary
This study introduces the Gait REcognition in the Wild (GREW) dataset, the first large-scale benchmark for unconstrained gait recognition. It also presents SPOSGait, a novel neural architecture search model achieving state-of-the-art results on multiple benchmarks.
Area of Science:
- Computer Vision
- Biometrics
- Machine Learning
Background:
- Existing gait recognition datasets are limited to controlled environments, hindering real-world application development.
- Cross-view gait recognition research requires more diverse and naturally challenging datasets.
Purpose of the Study:
- To introduce a large-scale, unconstrained gait recognition dataset named GREW (Gait REcognition in the Wild).
- To establish a strong baseline for unconstrained gait recognition research.
- To propose a novel neural architecture search (NAS) model for gait recognition.
Main Methods:
- Construction of the GREW dataset using natural videos from thousands of hours of streams across hundreds of cameras.
- Inclusion of 26,000 identities and 128,000 sequences with rich attributes, plus a 233,000-sequence distractor set.
- Development and application of SPOSGait, a Single Path One-Shot neural architecture search model tailored for gait recognition.
Main Results:
- The GREW dataset offers diverse view variations and natural challenges, surpassing limitations of controlled datasets.
- SPOSGait achieved state-of-the-art performance on CASIA-B, OU-MVLP, Gait3D, and the new GREW benchmark.
- SPOSGait demonstrated superior performance compared to existing gait recognition approaches.
Conclusions:
- The GREW benchmark is crucial for training and evaluating gait recognizers in unconstrained, real-world scenarios.
- SPOSGait represents a significant advancement as the first NAS-based gait recognition model, setting new performance standards.
