code modified as following:
| function | file | necessary |
|---|---|---|
| coco format annotation + ignore dataset support | + mmdet\datasets\cocofmt.py > mmdet/datasets/init.py |
Y |
| ScaleFactor=1.0 for ReSize; | > mmdet/datasets/pipelines/transforms.py:74,99,289 | Y |
| corner dataset support | > mmdet/datasets/pipelines/loading.py:64 | X |
| auto cut image and merge result durring inference | + mmdet/datasets/pipelines/rtest_time_aug.py > mmdet/datasets/pipelines/init.py > mmdet/core/bbox/transforms.py:bbox_mapping,bbox_mapping_back > mmdet/models/dense_heads/dense_test_mixins.py:192 mmdet/core/post_processing/merge_augs.py:69,102 > mmdet/models/detectors/two_stage.py:aug_test+tile_aug_test > mmdet/models/roi_heads/test_mixins.py:168,323 |
X |
| do final test | > do_final_eval：mmdet/core/evaluation/eval_hooks.py:10,40 | X |
| Scale Match | + mmdet/core/bbox/coder/bouding_box.py + mmdet/datasets/pipelines/scale_match.py > mmdet/datasets/pipelines/init.py |
X |
| stop while nan | > mmdet/apis/train.py:165,174; >${config}.py:add check=dict(stop_while_nan=True) |
X |
coco format annotation + ignore dataset support include：
- load coco format json annotation;
- filte out 'ignore' bbox while load annotation;
- TinyPerson evaluation support;
- Location evaluation support;
evaluate function invoke graph
tools/train.py:train_detector
-> mmdet/apis/train.py:eval_hook(,**cfg.get('evaluation'))
-> mmdet/core/evaluation/eval_hooks.py:evaluate
-> mmdet/core/evaluation/eval_hooks.py:dataset.evaluate(
results, logger=runner.logger, **self.eval_kwargs)
-> mmdet/datasets/cocofmt.py:def evaluate(self,
…,
cocofmt_kwargs={}):
cocoEval = COCOExpandEval(cocoGt, cocoDt, iou_type, **cocofmt_kwargs)