MimicGait: A Model Agnostic approach for Occluded Gait Recognition using Correlational Knowledge Distillation

Ayush Gupta, Rama Chellappa · 2025

Read MimicGait: A Model Agnostic approach for Occluded Gait Recognition using Correlational Knowledge Distillation by Ayush Gupta, Rama Chellappa (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-15666.txt.

Introduction

Ayush Gupta and Rama Chellappa teach gait recognition to work when part of the walker is hidden in this 2025 paper. Identifying people by their walk suits long distances, but outdoor scenes block the view with cars, crowds, and clutter. Most systems were trained on clean indoor footage and fail there.

Their method can wrap around existing recognition models. A mimic network learns from a teacher, guided by losses that capture how sequences relate to each other and how frames hang together within one walk. A side network estimates which parts of the body are visible and steers the learning toward the trustable parts. On three hard real-world datasets the approach held up under occlusion better than standard systems. The authors release their code for others building surveillance and search tools.

Catalog record

CREATORAyush Gupta, Rama Chellappa
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE73 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-15666

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MimicGait: A Model Agnostic approach for Occluded Gait Recognition using Correlational Knowledge Distillation Ayush Gupta, Rama Chellappa Abstract Gait recognition is an important biometric technique over large distances. State-of-the-art gait recognition systems perform very well in controlled environments at close range. Recently, there has been an increased interest in gait recognition in the wild prompted by the collection of outdoor, more challenging datasets containing variations in terms of illumination, pitch angles, and distances. An important problem in these environments is that of occlusion, where the subject is partially blocked from camera view. While important, this problem has received little attention. Thus, we propose MimicGait , a model-agnostic approach for gait recognition in the presence of occlusions. We train the network using a multi-instance correlational distillation loss to capture both inter-sequence and intra-sequence correlations in the occluded gait patterns of a subject, utilizing an auxiliary Visibility Estimation Network to guide the training of the proposed mimic network. We demonstrate the effectiveness of our approach on challenging real-world datasets like GREW, Gait3D and BRIAR. We release the code in https://github.com/Ayush-00/mimicgait . 1 Introduction Gait is an important biometric feature which can be used for identifying humans [ 28 ] , especially when the face is not visible. There has been significant progress in the field of Gait Recognition - the problem of recognizing subjects based on their walking pattern. Gait recognition can be performed by placing wearable sensors on subjects [ 10 , 17 ] , however, such methods require the subjects’ cooperation and are not scalable. With progress in computer vision, the popularity of gait recognition techniques using only vision-based modalities has risen significantly. As a result, gait has gained a unique importance among all biometric signatures as one of the few identifying characteristics in humans that can be captured effectively at a distance. Figure 1: Visualizations of the synthetic occlusions used in our experiments, taken from the GREW dataset. The original holistic video is shown in the top left. Middle occlusions are shown in the top right. The second row shows the same video with consistent synthetic occlusions. The bottom most row shows the same video with dynamic occlusions. The boundary of the moving occlusion patch is shown in red for visualization purposes only. There has been significant progress in the field of vision-based gait recognition [ 32 ] , with some methods achieving almost perfect scores on indoor controlled datasets [ 23 , 5 ] . With saturation in controlled scenarios, there has been an increased focus on outdoor, in the wild scenarios [ 45 , 2 ] . Such datasets pose much bigger challenges to gait recognition - due to large variations in viewpoint, altitude, clothing, background, illumination changes and occlusion. A deployed gait recognition system should be able to handle occlusion scenarios. There can be many types of occlusions; arising from an obstruction between the camera and the subject, or due to improper camera placement. Occlusion can be consistent, such as an elevated sidewalk blocking a subject’s feet for the entire sequence, or dynamic, such as another person or a stationary object temporarily blocking the subject of interest from view. With gait being recognized as a viable option for biometrics, it is important to address occlusion within gait recognition. Most current work on gait recognition does not address this problem specifically; the lack of a large-scale dataset focused on occlusion has resulted in slow progress in the area. Some works that focus on this problem simulate occlusions on indoor datasets [ 37 ] or work with small datasets [ 33 ] . Moreover, most current approaches assume an ideal occlusion scenario, where the subject is close to the camera and clearly visible. In a more practical in-the-wild scenario, the subject could be hundreds of meters away, and the camera may be situated at an altitude, and it is not easy to extend these approaches to such unconstrained data. [ 34 ] uses a generative network to synthesize complete gait sequences in the case of occlusion. However, it will be difficult to generate such sequences when the partial input is itself of low quality. Similarly, [ 37 ] uses an SMPL-based human mesh model to construct the gait signature, but the 3D structure of the body is not easy to recover from noisy data collected at a distance of several hundred metres, that too in the presence of occlusions. [ 13 ] generates occlusion aware features and inserts them inside the gait recognition network. However, it is limited by the assumption that the network can independently learn discriminative features through an occlusion detector, neglecting the potential correlations between occluded and visible body parts. To address these challenges, we propose MimicGait , a model-agnostic approach to generate discriminative gait features for subjects at range under occlusion using correlational knowledge distillation. We assume that temporal patterns which exist in the occluded sections of the subject are correlated with the observable motion in the gait sequence. We adopt a knowledge distillation approach to learn these correlations among the occluded and visible parts of the body, enabling the network to produce features closer to the ideal, holistic features. Building upon the work by [ 13 ] , we also utilize a Visibility Estimation Network (VEN) to introduce occlusion-relevant features into our method to enhance the prediction of these missing correlations. Our approach is model-agnostic and can be used to increase the performance of any state-of-the-art gait recognition method to extend it to occlusions. Lastly, being a vision-based method, it can work with noisy data captured at a distance. We demonstrate the performance of our proposed method on the publicly available GREW [ 45 ] and Gait3D [ 43 ] datasets. Additionally, we also evaluate our approach on the BRIAR [ 2 ] dataset, which includes variations in range, altitude, clothing and walking conditions. We also introduce some practical evaluation schemes for occluded gait recognition. We formalize the concept of Generalizability , introduced in [ 13 ] , relating model performance on unseen occlusions. We further introduce a concept of Adaptability , relating to how well the model can be trained/adapted to newer occlusions. This becomes useful where a certain type of occlusion is expected to occur frequently in deployment. Lastly, we propose another metric for occluded gait recognition called relative performance , RP, which measures the occluded recognition performance of the model relative to performance on ideal, holistic data. We show that it is a better metric to evaluate occluded performance. Together with generalizability, adaptability and RP, we perform an extensive analysis of our method and show that it outperforms previous works. In summary, our main contributions are • We propose MimicGait , a novel model-agnostic approach to generate robust gait features under occlusions in various conditions, viewpoints, and distances. • We demonstrate the utility of a multi-instance correlational knowledge distillation approach to learn the correlations between occluded and visible motion patterns across multiple gait instances of a subject. • We improve upon the auxiliary occlusion detector proposed in previous works and propose a Visibility Estimation Network to enhance gait recognition performance under occlusions. • We introduce the concepts of generalizability, adaptability, and a new metric RP for evaluating occluded gait recognition

Read the complete text