From Indoor To Outdoor: Unsupervised Domain Adaptive Gait Recognition

TL;DR

The paper tackles unsupervised domain adaptive gait recognition from indoor to outdoor, where labeled indoor data must generalize to unlabeled outdoor data. It proposes a three-stage framework—source-domain training, target-domain pseudo-label generation via clustering, and target-domain fine-tuning—augmented with uncertainty estimation to mitigate noisy pseudo labels. An uncertainty-regularized learning scheme weights samples and pairs by estimated uncertainty, and a new outdoor gait benchmark (DukeGait) with four cross-domain protocols is established. Experiments show the proposed method consistently outperforms state-of-the-art gait recognition and UDA-Re-ID baselines, demonstrating practical potential for real-world outdoor gait identification.

Abstract

Gait recognition is an important AI task, which has been progressed rapidly with the development of deep learning. However, existing learning based gait recognition methods mainly focus on the single domain, especially the constrained laboratory environment. In this paper, we study a new problem of unsupervised domain adaptive gait recognition (UDA-GR), that learns a gait identifier with supervised labels from the indoor scenes (source domain), and is applied to the outdoor wild scenes (target domain). For this purpose, we develop an uncertainty estimation and regularization based UDA-GR method. Specifically, we investigate the characteristic of gaits in the indoor and outdoor scenes, for estimating the gait sample uncertainty, which is used in the unsupervised fine-tuning on the target domain to alleviate the noises of the pseudo labels. We also establish a new benchmark for the proposed problem, experimental results on which show the effectiveness of the proposed method. We will release the benchmark and source code in this work to the public.

Figures (4)

• Figure 1: A comparison between the gait recognition task the indoor (laboratory) and outdoor (wild) environments. (a) An illustration of the shooting condition of an indoor datasets in takemura2018multi, where the cameras and walking courses of subjects are fixed. All samples have the same clean background. (b) A new outdoor gait recognition dataset built in this work, where the camera views and background are random and complex, belongings carrying and partial occlusions are also common.
• Figure 2: Framework of the proposed method for UDA-GR. In the source domain pre-training stage, we pre-train the model using the source domain samples with the given ground-truth labels. Then the clustering algorithm is applied on the features of unlabeled target domain samples to achieve pseudo labels. Finally, we fine-tune the model using the target domain data with the predicted pseudo labels. We estimate the sample uncertainty based on the discrepancy between the original and transformed gaits, which is used to modify the cross-entropy and triplet losses, to guide both the network pre-training and fine-tuning.
• Figure 3: Examples of common disturbed conditions, i.e., walking direction change, occlusion, and camera view change, in outdoor dataset. Each pair of RGB images (the first row) and the corresponding silhouettes (the second row) contain the same subject, and random flip, random erase, and random perspective transformation (the third row) are designed to simulate these conditions.
• Figure 4: Examples of pseudo label and uncertainty prediction results on DukeGait. The uncertainties are normalized by min-max scaling.