Real Time Object Detection for Traffic Based on Knowledge Distillation: 3rd Place Solution to Pair Competition

• Published in: 2020 IEEE International Conference on Multimedia & Expo Workshops (ICMEW) pp. 1 - 6
• Main Authors: Zhao, Yuan, Wang, Lyuwei, Hou, Luanxuan, Gan, Chunsheng, Huang, Zhipeng, Hu, Xu, Shen, Haifeng, Ye, Jieping
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.07.2020
• Subjects: Acceleration; Annotations; Computational modeling; Detectors; embedded system; Feature extraction; knowledge distillation; lightweight model; Object detection; Training
• DOI: 10.1109/ICMEW46912.2020.9105963

In practical applications, the purpose of object detection is to determine the target space position based on the image. At the same time, better performance is obtained under the premise of reducing the computational overhead. The dataset of the PAIR competition has the characteristics of imbalanced categories, low quality of images and inconsistent annotations. To address this issue, firstly we adopt an improved cross entropy loss function and data augmentations to rebalance the data distribution. Then the extra datasets are involved to neutralize the low images quality and annotation inconsistency issues. Secondly, this competition focuses on object detection on embedded device. So we apply knowledge distillation to fine-tune a lightweight detection model. Our detection model uses MobileNetV3 Small as backbone and SSDLite as detector head. In order to improve detection performance on small targets, FPNLite is included so that low-level features can be utilized. And we also apply TensorRT library to accelerate the inference procedure further. Eventually, our method achieves the 3 rd place in the final score list of competition as the fastest, lightest and the most computation economically solution. Our code will soon be open source.