YOLOv3

YOLOv3 implemented with distributed tensorflow including train and inference module.

Usage

1. generate anchors

put VOC 2007/2012 or custom files with VOC Annotation Format under data/ and run:

./run_prepare.sh -a

data meta info and anchor output as below:

max_width = 500, max_height = 500, max_box_per_image = 39
iteration 1: dists = 256748.369356
iteration 2: dists = 10246.3258411
...
... 
('\naverage IOU for', 9, 'anchors:', '0.68')
29,53, 61,94, 73,205, 134,119, 135,271, 218,329, 256,162, 342,247, 363,374

copy max_box_per_image and anchors into yolo_darknet53_voc.json or yourself config json file. max_width and max_height are used to be width and height when converting data to tfrecord.

2. convert data to tfrecord

convert VOC Annotation Format image and label into tfrecord under data/tfrecord/

./run_prepare.sh -t

3. train

confirm all parameters in the config file before begining to train, refering to yolo_darknet53_voc.json
download the darknet53.conv.74 or tiny.weights from yolo homepage and put them under conf/init/.
of course you can train with random initialization, for my limited experience, it could get similar mAP if your problem is not very complex like coco. I got very similar mAP on my other tasks execpt for convergence speed.

./run_train.sh -l

you must change the data path and model storage directory parameters and rewrite distributed_train() in run_train.sh, then replace -l with -r

4. convert checkpoint model to pb

convert checkpoint model to pb format for inference

./run_convert.sh -d

5. predict

predict the VOC2007 Test and output a single result file for every class.

./run_evaluate.sh -p

6. evaluate mAP

./run_evaluate.sh -e

Performance

Train Data	Test Data	Input Size	Model	Initialize	mAP	FLOPS
VOC2007+2012	VOC2007 Test	416*416	DarkNet53(YOLOv3)	darknet53.conv.74	0.782	65.86
VOC2007+2012	VOC2007 Test	416*416	YOLOFLY	tiny.weights	0.691	9.78

DarkNet53(YOLOV3)

aeroplane       AP = 0.869675
bicycle         AP = 0.858369
bird            AP = 0.745337
boat            AP = 0.714346
bottle          AP = 0.656200
bus             AP = 0.852547
car             AP = 0.877897
cat             AP = 0.860574
chair           AP = 0.652006
cow             AP = 0.777205
diningtable     AP = 0.750408
dog             AP = 0.815225
horse           AP = 0.868844
motorbike       AP = 0.848694
person          AP = 0.837868
pottedplant     AP = 0.536273
sheep           AP = 0.725731
sofa            AP = 0.754855
train           AP = 0.847557
tvmonitor       AP = 0.782320
mAP = 0.781596517434

Acknowledgements

Thanks to experiencor for his great job of keras-yolo3