Al_surface_defect_detection

This includes my code for Alibaba Tianchi competition: Al surface defect detection.

The competition is aimed at using computer vision techniques to help workers check whether their Al surface products have any defects such as spots, scratches and so on.

Season1(clasification) rank: 96/2972

What I found very useful:

• InceptionV4(pytorch)
• combine vote(similar to bagging)
• good iteration steps

Just so so:

• data augmentation(horizontal flip)
• Ensemble(Xception, Resnet50, InveptionV3)

Decrease my test acc

• All other augmentation. especially random rotation(0~8, the larger angle, the worse acc))

Season2(localization) rank：10/2972

What I found very useful:

• FasterRcnn&FPN(detectron)
• Larger resize size(960 for maskrcnn, 800 for FasterRcnn)
• bbox vote
• Adam instead of SGD
• lower the thresh

Just so so:

• Mask-Rcnn(keras tf)
• YoloV3
• FasterRcnn(tf)
• Emsemble(FPN, faster-rcnn, mask-rcnn)
• Soft-nms(since few defects have overlap)
• bbox-vote strategy(ID, AVG, IOU_AVG)
• Delete mini batch(since spots are super small)
• Data augmentation( train&test scales, flip, small rotation)

Decrease my test acc:

• my own bbox vote( similar to softer-nms, a combination of iou and confidence)
• bbox combination ( similar to this from kaggle big bowl 2018)
• Use larger size(1920x2560), more data augmentation(5 scales etc.)...

What I didn't have time to try:

• SNIPER
• Cascade-rcnn
• maskrcnn(X152 backbone)

Possible useful links:

• Official discussion forum
• DCN github link
• 1st place's article sharing