MaskRCNN

Implementation of mask r-cnn for practice

Week 0

Surveyed Related Papers

https://docs.google.com/presentation/d/1OyoITuqV_OD0Y37HTnhrMvYxfm3jYLLiMSrDXHctu_E/edit?usp=sharing

Week 1

TODO:Implement dataloader of COCO, backbone models

02.14 implemented coco dataloader with img, bboxes

GT_bboxes visualized

Week 2

TODO:

• skeleton list up, outline
• dataloader augmentation(detectron2 ref.) +bbox transformation

02.21 Used Gyeongsik's codes as a reference and created baseline skeletons based on them

02.24

• Image preprocessing Implemented(resize to 1024**2, zeropadding, transformations)

original image

transformed image (1024x1024 zeropadded aspect_remained)

bbox also transformed properly under image preprocessing

• FPN and Anchor Generation code Implemented

anchors under various feature levels

whole_anchors

02.25 Implemented IoU(simple) function

Week3

TODO:

• RPN bbox_reg, objectiveness head implementation

03.02 Implemented and Visualized Pos&Neg anchor(IoU > 0.5) 03.03 Implemented IoU tensor broadcasting instead of for loops

positive anchors under various feature levels

03.05 implemented anchor_labeling, subsampling codes / visualized pos/neg windows

Week4

03.10 Implemented rpn losses, transform_delta, RPN under train

Week5

03.13 Tested rpn on one image

03.15 Implemented NMS function 03.21 Trained RPN, visualized qualitative results

TODO: ROI align