[Add] Support VIS evaluation for YouTube-VIS task. #501
Conversation
Codecov Report
@@ Coverage Diff @@
## master #501 +/- ##
==========================================
- Coverage 73.35% 73.06% -0.30%
==========================================
Files 126 129 +3
Lines 7441 8055 +614
Branches 1395 1562 +167
==========================================
+ Hits 5458 5885 +427
- Misses 1555 1675 +120
- Partials 428 495 +67
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
|
I need to submit two JSON files for unit testing, but they can't pass Pre-commit. How should I solve this problem? |
|
The overall structure needs to be adjusted, we should try to reuse the current code and only add some necessary functions. Some suggestions are as the following:
|
| for idx in instance_info: | ||
| cur_video_len = len_video[instance_info[idx][0]['video_id']] | ||
| len_videos[img_info['video_id']] = max(1, img_info['frame_id'] + 1) | ||
| for ins_id in instance_infos: |
There was a problem hiding this comment.
The last review has mentioned! Too many instance_infos[ins_id] in this piece of code, we could set instance_infos = instances_infos[ins_id], and use instance_infos rather than instance_infos[ins_id] in the following code
| bbox[frame_id[ann_info['image_id']]] = ann_info['bbox'] | ||
| area[frame_id[ann_info['image_id']]] = ann_info['area'] | ||
| for ann_info in instance_infos[ins_id]: | ||
| segm[frame_id_mapping[ |
There was a problem hiding this comment.
The same problems as the above. Too many repeated use of dictionary indexes this piece of code.
frame_id = frame_id_mapping[ann_info['image_id']], and only use frame_id below.
| len_video[img_info['video_id']] = max(1, img_info['frame_id'] + 1) | ||
| for idx in instance_info: | ||
| cur_video_len = len_video[instance_info[idx][0]['video_id']] | ||
| len_videos[img_info['video_id']] = max(1, img_info['frame_id'] + 1) |
There was a problem hiding this comment.
| len_videos[img_info['video_id']] = max(1, img_info['frame_id'] + 1) | |
| video_id = img_info['video_id'] | |
| len_videos[video_id] = len_videos.get(video_id, 0) + 1 |
| instance_info = defaultdict(list) | ||
| frame_id = defaultdict(list) | ||
| len_video = defaultdict(list) | ||
| instance_infos = defaultdict(list) |
There was a problem hiding this comment.
Deprecate the old code, we could use the coco api to realize this function as the following:
VIS = defaultdict(list)
the 'categories' and 'videos' has already loaded in this dataset, don't need to reload it.
for ins_id, img_ids in self.instancesToImgs.items():
instance_infos = []
for img_id in img_ids:
for ann in self.coco.imgToAnns[img_id]:
if ann['instance_id'] == instance_id:
add some code to load ann information of this instance
VIS['annotations'].append(instance_infos)
|
Deprecate the old code of VIS = defaultdict(list)
the 'categories' and 'videos' has already loaded in this dataset, don't need to reload `self.ann_file`.
for ins_id, img_ids in self.coco.instancesToImgs.items():
instance_infos = []
for img_id in img_ids:
for ann in self.coco.imgToAnns[img_id]:
if ann['instance_id'] == instance_id:
add some code to load ann information of this instance
VIS['annotations'].append(instance_infos)Please check the above pseudo-code to make it run rightly. |
update youtube_vis
There are three main modifications here.
mmtrack/core/evaluation/eval_vis.pywidthandheightinfo to the key ofvideo.evaluateinmmtrack/datasets/youtube_vis_dataset.py.For the first part, the JSON format used for training is converted into the native JSON format of YouTube-VIS for calculation.