QGait: Toward Accurate Quantization for Gait Recognition

TL;DR

Gait recognition faces efficiency constraints for edge deployment, and naive quantization can degrade performance on binarized silhouette inputs. The authors propose QGait, which combines a differentiable soft quantizer with a two-stage training strategy and Inter-class Distance-guided Calibration (IDC) to preserve inter-class separation under low-bit quantization. Empirical results across Gait3D, GREW, and OUMVLP demonstrate that 8-bit quantization can match full-precision accuracy while 4-bit quantization benefits from IDC and staged training, enabling substantial speedups with minimal loss. This work enables practical, high-accuracy gait recognition on resource-constrained devices and offers a general framework for quantization-aware appearance-based biometric systems.

Abstract

Existing deep learning methods have made significant progress in gait recognition. Quantization can facilitate the application of gait models as a model-agnostic general compression technique. Typically, appearance-based models binarize inputs into silhouette sequences. However, mainstream quantization methods prioritize minimizing task loss over quantization error, which is detrimental to gait recognition with binarized inputs. To address this, we propose a differentiable soft quantizer, which better simulates the gradient of the round function during backpropagation. This enables the network to learn from subtle input perturbations. However, our theoretical analysis and empirical studies reveal that directly applying the soft quantizer can hinder network convergence. We addressed this issue by adopting a two-stage training strategy, introducing a soft quantizer during the fine-tuning phase. However, in the first stage of training, we observed a significant change in the output distribution of different samples in the feature space compared to the full-precision network. It is this change that led to a loss in performance. Based on this, we propose an Inter-class Distance-guided Calibration (IDC) strategy to preserve the relative distance between the embeddings of samples with different labels. Extensive experiments validate the effectiveness of our approach, demonstrating state-of-the-art accuracy across various settings and datasets.

Figures (6)

• Figure 1: We modified the video-based All-in-One Gait method Fan2022OpenGaitRG for real-time video stream processing. Tracking and segmentation were incorporated into the frame loop to collect gait sequences. We defined a sequence period as 30 frames and found that the gait recognition component accounted for a significant portion of the overhead. The tracking model uses ByteTrack Zhang2021ByteTrackMT, and the segmentation model employs HumanSeg-V2-mobile with PaddleInference liu2021paddleseg. By quantizing the gait recognition module, we achieved over a twofold acceleration.
• Figure 2: Parameters and performance comparison. QGait outperforms state-of-the-art methods with a relatively small number of parameters. Comparisons are performed on the Gait3D dataset Zheng2022GaitRI.
• Figure 3: The pipeline of our proposed QGait.
• Figure 4: (a) Visualization of the round function along with the graph of ${\theta}_{k}(x)$ used for its approximation when $k$ takes different values. (b) (c) Visualization of $\mathbb{E}^2\left[ \nabla G(z)\right]$ and $\mathbb{E}[\|\nabla G(z)\|^2]$.
• Figure 5: Visualization using t-SNE of the full-precision network, 4-bit network without IDC and 4-bit network with IDC. The results indicate that IDC can significantly mitigate the changes in inter-class distances.