This is the implementation of "Enriching Variety of Layer-wise Learning Information by Gradient Combination" using Darknet framwork.
Our paper will be appear in 2019 ICCV Workshop on Low-Power Computer Vision.
For installing Darknet framework, you can refer to darknet(pjreddie) or darknet(AlexeyAB).
We provide YOLO-v3-tiny-PRN cfg file and COCO pre-trained model. You can use provided files to get following results on COCO test-dev set:
Model | mAP@0.5 | BFLOPs | # Parameter | GPU FPS | CPU FPS |
---|---|---|---|---|---|
YOLO-v3-tiny | 33.1 | 5.571 | 8.86M | 300 | 8 |
YOLO-v3-tiny-PRN | 33.1 | 3.467 | 4.95M | 370 | 13 |
We also provide cfg file and COCO pre-trained model for morden backbone EfficientNet_b0. For training this model, you should install darknet(AlexeyAB).
Model | Size | mAP@0.5 | BFLOPs |
---|---|---|---|
EfficientNet_b0-PRN | 416x416 | 45.5 | 3.730 |
EfficientNet_b0-PRN | 320x320 | 41.0 | 2.208 |