Xiangyu He, Ke Cheng, Qiang Chen, Qinghao Hu, Peisong Wang, Jian Cheng
Long-range dependencies modeling, widely used in capturing spatiotemporal correlation, has shown to be effective in CNN dominated computer vision tasks. Yet neither stacks of convolutional operations to enlarge receptive fields nor recent nonlocal modules is computationally efficient. In this paper, we present a generic family of lightweight global descriptors for modeling the interactions between positions across different dimensions (e.g., channels, frames). This descriptor enables subsequent convolutions to access the informative global features with negligible computational complexity and parameters. Benchmark experiments show that the proposed method can complete state-of-the-art long-range mechanisms with a significant reduction in extra computing cost. Code available at https://github.com/HolmesShuan/Compact-Global-Descriptor.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Object Detection | COCO test-dev | box mAP | 37.9 | Faster R-CNN + FPN + CGD |
| Image Classification | ImageNet | GFLOPs | 1.198 | MobileNet-224 (CGD) |
| 3D | COCO test-dev | box mAP | 37.9 | Faster R-CNN + FPN + CGD |
| 2D Classification | COCO test-dev | box mAP | 37.9 | Faster R-CNN + FPN + CGD |
| 2D Object Detection | COCO test-dev | box mAP | 37.9 | Faster R-CNN + FPN + CGD |
| 16k | COCO test-dev | box mAP | 37.9 | Faster R-CNN + FPN + CGD |