Home

Awesome

Deformable ConvNets is initially described in an arxiv tech report.

R-FCN is initially described in a NIPS 2016 paper.

Soft-NMS is initially described in an arxiv tech report.

Our goal was to test Soft-NMS with a state-of-the-art detector, so Deformable-R-FCN was trained on 800x1200 size images with 15 anchors. Multi-Scale testing was also added with 6 scales. Union of all boxes at each scale was computed before performing NMS. Please note that the repository does not include the scripts for multi-scale testing as I just cache the boxes for each different scale and do NMS separately. The scales used in multi-scale testing were as follows, [(480, 800), (576,900), (688, 1100), (800,1200), (1200, 1600), (1400, 2000)].

The trained model can be downloaded from here.

<sub>training data</sub><sub>testing data</sub><sub>mAP</sub><sub>mAP@0.5</sub><sub>mAP@0.75</sub><sub>mAP@S</sub><sub>mAP@M</sub><sub>mAP@L</sub><sub>Recall</sub>
<sub>Baseline D-R-FCN</sub><sub>coco trainval</sub><sub>coco test-dev</sub>35.756.838.315.238.851.5
<sub>D-R-FCN, ResNet-v1-101, NMS</sub><sub>coco trainval</sub><sub>coco test-dev</sub>37.459.640.217.840.651.448.3
<sub>D-R-FCN, ResNet-v1-101, SNMS</sub><sub>coco trainval</sub><sub>coco test-dev</sub>38.460.141.618.541.652.553.8
<sub>D-R-FCN, ResNet-v1-101, MST, NMS</sub><sub>coco trainval</sub><sub>coco test-dev</sub>39.862.443.322.642.352.252.9
<sub>D-R-FCN, ResNet-v1-101, MST, SNMS</sub><sub>coco trainval</sub><sub>coco test-dev</sub>40.962.845.023.343.653.360.4