swoook / yolox Goto Github PK

🚀 Feature

• TTA including multi-scales, reflection, and so on

Motivation & Examples

• See the example of TTA for YOLOv5 at here

🚀 Feature

• I'd like to use some alternatives to Non-Maximum Suppression (NMS) below:

1. soft-NMS
2. Non-Maximum Weighted (NMW)
3. Weighted Boxes Fusion (WBF)

Motivation & Examples

• NMS is one of the most conventional techniques to eliminate redundant bounding boxes
• But some techniques show better results than NMS:

1. soft-NMS
2. Non-Maximum Weighted (NMW)
3. Weighted Boxes Fusion (WBF)

• Soft-NMS surpasses NMS on PASCAL VOC 2007 and MS-COCO for R-FCN and Faster-RCNN

🚀 Feature

• I'd like to try vertical flip for data augmentation

Motivation & Examples

• I've seen many works don't borrow vertical flip for data augmentation
• Then, is it inappropriate to apply vertical flip to data augmentation in object detection?
• It depends on type of domains or datasets
• I benefited from it when I studied object detection in x-ray images