TL;DR
This paper introduces WaveLoss, a novel loss function for deep gait recognition that dynamically adjusts its focus on different body parts based on their difficulty during training. By adaptively learning from challenging gait sequences, WaveLoss significantly improves recognition accuracy across various datasets and backbone architectures.
What Problem Does This Paper Address?
Deep learning has revolutionized gait recognition, but most research has focused on improving network architectures or input modalities, largely overlooking the crucial role of loss functions. Existing loss functions, often adapted from face recognition, struggle with the unique challenges of gait data, such as the need to consider different body parts independently and the difficulty of distinguishing between low-quality and genuinely difficult samples. This can lead to slow convergence, susceptibility to noise, and suboptimal performance, especially in complex, real-world scenarios.
Methods
The core innovation is WaveLoss, which combines and refines existing fusion strategies for part-level gait features.
- Norm-Fusion: This acts as an intermediate between "pre-fusion" (averaging all part features before loss calculation) and "post-fusion" (calculating loss for each part individually and then averaging). Norm-fusion uses a weighted sum of part losses, where the weights are dynamically adjusted. This allows the model to balance the benefits of both pre-fusion (robustness to noise) and post-fusion (focus on difficult parts).
- Dynamic Learning Process: To further enhance adaptability, the weighting exponent ($l$) in Norm-Fusion is dynamically adjusted throughout training. This allows the model to shift its focus from easier samples early in training to more challenging ones later, mimicking curriculum learning and hard sample mining. The exponent $l$ follows a sinusoidal pattern, enabling the model to repeatedly adapt its learning preference.
- Gradient Truncation and Normalization: To mitigate the instability introduced by the dynamic exponent, the authors employ gradient truncation to smooth out extreme gradient values and gradient normalization to calibrate the step size. This prevents gradient explosion/vanishing and helps avoid local optima, ensuring stable convergence.
Datasets
The proposed WaveLoss was evaluated on three popular gait recognition datasets:
- CASIA-B: An indoor dataset widely used for gait analysis, featuring data from 124 subjects under normal walking, walking with a bag, and walking in coats conditions, with varying viewing angles.
- OUMVLP: A large-scale indoor dataset with data from over 10,000 subjects, captured under different conditions.
- Gait3D: An outdoor dataset featuring data from 4,000 subjects captured using 39 cameras, notable for its inclusion of 3D Skinned Multi-Person Linear (SMPL) models.
Key Insights and Ideas
- Part-level loss is critical for gait recognition: Unlike face recognition, gait inherently involves distinct body parts (e.g., legs, torso) that contribute differently to recognition. Treating these parts equally or simply averaging them can be suboptimal.
- Adaptive weighting of part losses is key: The ability to dynamically adjust the importance of different parts based on their learning difficulty is crucial for robust gait recognition.
- Dynamic learning schedules can be beneficial: Mimicking curriculum learning by gradually increasing the focus on harder samples during training can significantly improve model performance.
- Stabilizing dynamic loss functions is essential: The proposed gradient truncation and normalization techniques are vital for making dynamic loss functions practical and preventing training instability.
Key Results
WaveLoss consistently achieved state-of-the-art (SOTA) performance across all tested datasets:
- On OUMVLP, WaveLoss achieved SOTA accuracy, demonstrating its effectiveness in cross-view feature extraction.
- On CASIA-B, WaveLoss achieved higher accuracy on challenging samples (CL), indicating its superior ability to increase intra-class similarity and reduce inter-class similarity.
- On Gait3D, WaveLoss showed significant improvements, particularly on this low-quality outdoor dataset, highlighting its adaptability to varying part quality.
- The ablation studies confirmed the effectiveness of both Norm-Fusion and the Dynamic Learning Process, showing improvements in accuracy and performance on difficult samples.
Limitations and Future Work
- Limitations: The paper acknowledges that while Norm-Fusion offers advantages, it's still challenging to simultaneously capture the benefits of both pre-fusion and post-fusion perfectly. The dynamic learning process, while effective, introduces sensitivity to the exponent $l$, which is addressed by gradient stabilization techniques.
- Future Work: The authors hope their proposed loss function will inspire further research into part-level loss fusion for gait recognition, suggesting it provides a new perspective for the field.
Summary updated