Pruned Lightweight Encoders for Computer Vision
This work addresses latency issues in autonomous driving or drone control by improving compression for offloading inference, but it is incremental as it uses existing encoders and retraining without architectural changes.
The paper tackles the problem of latency-critical computer vision systems by using lightweight encoders with constant bitrate and pruned configurations (ASTC and JPEG XS) to reduce encoding speed, and recovers distortion through retraining neural networks with compressed data, achieving reductions in accuracy degradation to 4.9-5.0 pp for classification and 4.4-4.0 pp for segmentation with ASTC, and 2.7-2.3 pp for segmentation with JPEG XS, while making the ASTC encoder 2.3x faster than JPEG.
Latency-critical computer vision systems, such as autonomous driving or drone control, require fast image or video compression when offloading neural network inference to a remote computer. To ensure low latency on a near-sensor edge device, we propose the use of lightweight encoders with constant bitrate and pruned encoding configurations, namely, ASTC and JPEG XS. Pruning introduces significant distortion which we show can be recovered by retraining the neural network with compressed data after decompression. Such an approach does not modify the network architecture or require coding format modifications. By retraining with compressed datasets, we reduced the classification accuracy and segmentation mean intersection over union (mIoU) degradation due to ASTC compression to 4.9-5.0 percentage points (pp) and 4.4-4.0 pp, respectively. With the same method, the mIoU lost due to JPEG XS compression at the main profile was restored to 2.7-2.3 pp. In terms of encoding speed, our ASTC encoder implementation is 2.3x faster than JPEG. Even though the JPEG XS reference encoder requires optimizations to reach low latency, we showed that disabling significance flag coding saves 22-23% of encoding time at the cost of 0.4-0.3 mIoU after retraining.