Compressing the Multiobject Tracking Model via Knowledge Distillation

Recent multiobject tracking (MOT) methods usually use very deep neural networks to achieve competitive accuracy, which inevitably results in degraded inference speed. To strike a better balance between tracking accuracy and speed, in this work, we propose to compress the MOT model via knowledge distillation (KD), enabling the more lightweight student model to obtain similar performance as the teacher model. Nonetheless, despite KD has been well studied for simpler tasks such as image classification, the complexity of MOT poses new challenges because the MOT model is more sensitive to foreground information than the classification model. To deal with that, we first propose attention-guided feature distillation, which focuses the student model on the crucial region (foreground and the region with strong discrepancy against itself) of the teacher's feature map. Moreover, we propose foreground mask, which leverages the knowledge from the teacher model to filter out the low-quality soft labels from the background, thereby reducing their negative effects for distillation. Evaluations on several benchmarks demonstrate that the proposed KD method can make the student network achieve leading performance, meanwhile running faster than the teacher network 20.0%-27.4% and reducing the parameters 28.5%-87.1%. To the best of our knowledge, this is the first work to compress the MOT model via KD.