CPKD: Channel and Position-Wise Knowledge Distillation for Segmentation of Road Negative Obstacles

Segmentation of road negative obstacles is important for the safety of autonomous vehicles. Many multi- modal fusion networks have been proposed for this task. They have achieved acceptable performance. However, most of them are heavyweight, making them hard to run real- timely, especially when working with high-resolution images. To address this issue, we propose a channel and position- wise knowledge distillation framework to train a lightweight student to achieve comparable accuracy and better efficiency. Specifically, we introduce a downsampling layer at the beginning of the student network to reduce the input data size to the student network, and introduce an upsampling layer at the end to restore the resolution. We propose a channel and position - wise distillation module to transfer knowledge between different sizes of feature maps. In addition, we release an RGB- Depth dataset for negative-obstacle segmentation. Experimental results demonstrate the effectiveness of our proposed method. Our code and dataset are available at: https://github.com/lab-sun/CPKD.