zcc720 / Tensorflow_YOLOv1

You Only Look Once: Unified, Real-Time Object Detection

You Only Look Once: Unified, Real-Time Object Detection

# Summary

1. end-to-end
2. extremely fast (base YOLO : 45 FPS, Fast YOLO : 155 FPS, other SOTA 7~20 FPS)
3. YOLOv1-VGG16 mAP : 66.4% vs FasterRCNN mAP : 73.2% (Pascal VOC 2007 Test)
4. small objects do not detect well.
5. each grid cell only predicts two boxes and can only have one class.

# Difference from Paper

1. GoogleNet -> ResNetv2-50, VGG16, Inception ResNet v2
2. Removed the responsible part.

# Results (Pascal VOC 2007 Test AP@50)

Class	(self) YOLOv1-VGG16	(self) YOLOv1-ResNetv2-50
aeroplane	75.23%	68.38%
bicycle	78.30%	70.28%
bird	67.54%	63.91%
boat	55.93%	49.64%
bottle	64.49%	57.35%
bus	78.75%	81.63%
car	77.83%	71.31%
cat	79.14%	78.95%
chair	63.65%	56.38%
cow	73.61%	72.50%
diningtable	68.76%	66.00%
dog	80.46%	76.45%
horse	79.26%	70.27%
motorbike	80.54%	73.30%
person	75.69%	70.37%
pottedplant	54.15%	56.02%
sheep	79.67%	75.43%
sofa	71.64%	67.10%
train	74.86%	76.29%
tvmonitor	76.21%	73.99%

Model	Train	mAP@50
(paper) Faster-RCNN VGG16	Pascal VOC 2007 + 2012	73.2%
(paper) YOLOv1 VGG16	Pascal VOC 2007 + 2012	66.4%
(self) YOLOv1 VGG16	Pascal VOC 2007 + 2012	72.78%
(self) YOLOv1 ResNet-v2-50	Pascal VOC 2007 + 2012	68.78%
(self) YOLOv1 Inception-ResNet-v2	Pascal VOC 2007 + 2012	69.05%

# Tensorboard

1. xy loss

2. wh loss

3. obj loss

4. noobj loss

5. class loss

# Test Samples

# Reference

• You Only Look Once: Unified, Real-Time Object Detection [Paper]
• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks [Paper]