Body Feature Filter for Occluded Person Re-Identification

Occluded person re-identification attracted widespread attention in recent years since it is more consistent with the real-world scenarios compared with other tasks of person reidentification. Researchers proposed a lot of creative works to tackle this task and achieved inspiring progress. However, most of these works severely relied on extra cues from some pretrained models, such as pose estimator, which made the result of their models sensitive to the estimation error in the extra cues. Moreover, they utilized pose estimator in both training and inference stage, which slowed down the inference speed a lot. In this paper, we propose a novel method named body feature filter (BFF), which is robust to inaccurate pose information from pose estimator by simply transferring score maps of key-points to body part label. Our BFF helps the model filter out noisy information from the occluded regions or the background and focus on the body features of person images. Our method consists of three steps in the training stage. In the first step, we utilize score maps of key-points from pose estimator to get local feature vectors. Then we calculate similarity map in which the value of each location is the similarity score between the local feature vector and the global feature of this location. Finally, we add similarity maps related to the same body part together to generate body part label. In the second step, we design a simple network, BFF, to learn the information from body part label. It generates body part masks to guide the model in knowing where is noisy feature to suppress and where is body part feature to enhance. In the third step, we add a module named body feature refiner (BFR) to further refine the body feature. In the inference stage, we only use body feature filter and body feature refiner without any cues from pose estimator to get final features for person retrieval. The simple architecture of our model makes inference speed faster than that of most other works. Besides, our BFF can be utilized in many person re-identification models, which uses pose estimator. Experiments show the effectiveness of our proposed method.