ByteDance launches a multi-object tracking library named “ByteTrack”
The library aims to solve the problem of detecting objects with low detection scores, for example occluded objects that are simply thrown away, and bring in significant missing true objects and fragmented trajectories. ByteTrack uses a simple, efficient and generic association method, tracking by associating each detection box instead of just the ones with the best score. For low scoring detection boxes, it uses their similarities to tracklets to retrieve real objects and filter out background detections.
ByteTrack uses BYTE technology, which is different from traditional methods, which only keep the high score detection boxes. BYTE keeps each detection box and separates them into high-scoring and low-scoring ones. It first associates the high score detection boxes with the tracklets. Some tracklets don’t match because they don’t match a proper high score detection box, which usually happens when an occlusion, motion blur, or size change occurs. Then, it combines the low score detection boxes and these unmatched tracklets to retrieve the items in the low score detection boxes and filter the background simultaneously.
The input to BYTE is a V video sequence, with an object detector and the Kalman filter. ByteTrack is equipped with a high performance detector named YOLOX, as well as the BYTE association method. YOLOX runs YOLO series detectors in a dockless manner and uses other advanced detection techniques including decoupled heads, large data boosts, such as Mosaic and Mixup with efficient SimOTA tag assignment strategy , to obtain peak performance on object detection
ByteTrack was evaluated on half of the MOT17 validation set using different combinations of training data. When using only half of the MOT17 drive assembly, the performance reaches 75.8 MOTA, outperforming most methods. This is because it uses strong augmentations such as Mosaic and Mixup. By adding CrowdHuman, Cityperson and ETHZ for training, we can achieve 76.7 MOTA and 79.7 IDF1.
ByteTrack is very robust to occlusion for its precise detection performance and aid in pairing low score detection boxes. The model also highlights the optimal use of detection results to improve multi-object tracking. The research team hopes that ByteTrack’s high precision, fast speed and simplicity can make it attractive and efficient in real-world applications.
Subscribe to our newsletter
Receive the latest updates and relevant offers by sharing your email.