Gaussian YOLOv3

This is an implementation of Gaussian YOLOv3 on Keras and Tensorflow. The project is based on qqwweee/keras-yolo3 and fizyr/keras-retinanet. Thanks for their hard work. The official implementation is jwchoi384/Gaussian_YOLOv3

As in the official implementation, there is a hyper parameter sigma_const. I get best result on Pascal VOC when set it to be 0.3. The best result (mAP₅₀ 0.8535) is a little better than the one (mAP₅₀ 0.8500) I got by original YOLOv3. I am not sure whether the improvement is due to random training or benefits from Gaussian logic.

Test

I trained on Pascal VOC2012 trainval.txt + Pascal VOC2007 train.txt, and validated on Pascal VOC2007 val.txt. There are 14041 images for training and 2510 images for validation.
The best evaluation result (score_threshold=0.01, mAP₅₀, image_size=416, sigma_const=0.3) on VOC2007 test is 0.8535.
Pretrained official yolo weights on COCO and gaussian yolo weights on Pascal VOC are here.
baidu netdisk extract code: qgnd google driver
python3 yolo/inference.py to test your image by specifying image path and model path there.

Train

build dataset (Pascal VOC, other types please refer to fizyr/keras-retinanet)

Download VOC2007 and VOC2012, copy all image files from VOC2007 to VOC2012.
Append VOC2007 train.txt to VOC2012 trainval.txt.
Overwrite VOC2012 val.txt by VOC2007 val.txt.

train

STEP1: python3 train.py --freeze-body yolo --gpu 0 --batch-size 32 --random-transform --compute-val-loss pascal datasets/VOC2012 to start training with lr=1e-3 then stop when val mAP keep dropping.
STEP2: python3 train.py --snapshot <xxx> --freeze-body darknet --gpu 0 --batch-size 32 --random-transform --compute-val-loss pascal datasets/VOC2012 to start training with lr=1e-4 then top when val mAP keep dropping.
STEP3: python3 train.py --snapshot <xxx> --freeze-body none --gpu 0 --batch-size 32 --random-transform --compute-val-loss pascal datasets/VOC2012 to start training with lr=1e-5 and then set lr=1e-6 when val mAP when keep dropping.

Evaluate