search

jbwang1997 / OBBDetection

OBBDetection is an oriented object detection library, which is based on MMdetection.
Apache License 2.0
520 stars 111 forks source link

Annotations vs Bounding Box definition #115

Open chandlerbing65nm opened 2 years ago

chandlerbing65nm commented 2 years ago

The DOTA dataset annotations are specified in corner positions (x1, y1, x2, y2, x3, y3, x4, y4) but why did you define your bounding box in center+height+width (x,y,w,h) position?

Does it mean that you converted the annotations from corner positions to center+height+width? If so, where in the code can I find these?

jbwang1997 commented 2 years ago

I load (x1, y1, x2, y2, x3, y3, x4, y4) as a mask and convert it into (x, y, w, h, t) by the Mask2OBB pipeline.

chandlerbing65nm commented 2 years ago

I load (x1, y1, x2, y2, x3, y3, x4, y4) as a mask and convert it into (x, y, w, h, t) by the Mask2OBB pipeline.

@jbwang1997 Can I ask where specifically the conversion code (Mask2OBB) of Faster-RCNN and RetinaNet detectors are located? I have tried searching for it but I cannot find it. The part where you call [Mask2OBB] function.

chandlerbing65nm commented 2 years ago

I load (x1, y1, x2, y2, x3, y3, x4, y4) as a mask and convert it into (x, y, w, h, t) by the Mask2OBB pipeline.

https://github.com/jbwang1997/OBBDetection/blob/11d1c6f7a085659389e25a078e1438cf37ddd31e/configs/obb/_base_/datasets/dota.py#L18

That is the one right?