implementing data preparation of RCNN paper from scratch as follow:
- download VOC dataset .tar file
- extract dataset file
- for each image :
-
- parse its XML file to extract bounding box (bb) and corresponding class
-
- perform the selective search to get the bb proposals
-
- for each bb resulted from step (5):
-
-
- perform NMS with IoU of .5 with groundtruth else assign background label for it
-
-
-
- calculate the groundtruth bb offset (used for bb regression )
-
- return computed bb with its assigned label from step (7)