PASCAL VOC2007 Test
| Model | Original | ChainerCV (weight conversion) |
|---|---|---|
| YOLOv2 | 75.8 % * | 75.8 % |
| YOLOv3 | 80.2 % | 80.2 % |
Scores are mean Average Precision (mAP) with PASCAL VOC2007 metric.
*: Although the original paper [1] reports 76.8 %, the darknet implementation and the provided weights achieved the lower score. Similar issue is reported here.
Detect objects in an given image. This demo downloads Pascal VOC pretrained model automatically if a pretrained model path is not given.
$ python demo.py [--model yolo_v2|yolo_v3] [--gpu <gpu>] [--pretrained-model <model_path>] <image>.jpg
Convert *.weights to *.npz. YOLOv2 and YOLOv3 are supported.
Note that the number of classes should be specified if it is not 80 (the number of classes in COCO).
$ python darknet2npz [--model yolo_v2|yolo_v3] [--n-fg-class <#fg_class>] <source>.weights <target>.npz
The evaluation can be conducted using chainercv/examples/detection/eval_voc07.py.
- Joseph Redmon et al. "YOLO9000: Better, Faster, Stronger" CVPR 2017.
- Joseph Redmon et al. "YOLOv3: An Incremental Improvement" arXiv 2018.