Backbone | Training | Testing | Input | FPS | mAP | [email protected] | [email protected] | mAP@S | mAP@M | mAP@L | |
---|---|---|---|---|---|---|---|---|---|---|---|
Ours_384 | VGG-16 | trainval35k | test-dev | 384×384 | 62.5 | 35.2 | 55.9 | 38.1 | 17.7 | 38.2 | 48.3 |
Ours_512 | VGG-16 | trainval35k | test-dev | 512×512 | 38.5 | 37.6 | 58.7 | 41.0 | 21.0 | 40.4 | 49.5 |
Ours_384 | ResNet-101 | trainval35k | test-dev | 384×384 | 40.0 | 36.9 | 57.0 | 40.2 | 16.4 | 40.4 | 53.3 |
Ours_512 | ResNet-101 | trainval35k | test-dev | 512×512 | 28.6 | 40.0 | 60.8 | 43.8 | 21.3 | 43.9 | 53.9 |
- Results of our models on minival set are 34.8, 37.6, 36.7, 39.6 respectively.
- The above results are obtained by single-scale training and testing without adopting scale-jitter training and inference augmentation (e.g., multi-scale, flip, voting)
- The runtime are measured on our local machine with single NVIDIA GTX 1080 Ti, i7-6850k CPU, pyTorch 0.4.1, CUDA 9.0 and cuDNN v7.0. Different configures may induce various runtime.
- We have updated the runtime of Ours_384 with ResNet-101 since the data of last commit 42ab085 was obtained when the gpu was busy.
- V means the backbone of VGG-16 and R is ResNet-101.
- Release the evaluation codes and models
- Release the training codes