Add Instance Segmentation Transforms#836
Conversation
yutong-xiang-97
changed the title
|
/review |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b7caf1ca53
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if isinstance(bboxes_out, list): | ||
| bboxes_out = np.array(bboxes_out) |
There was a problem hiding this comment.
When an image has no instances (which the dataset allows by default) or an augmentation drops every box, albumentations can return bboxes as an empty list; np.array([]) produces shape (0,), so the dataset returns a box tensor with no second dimension. The downstream matchers/criteria expect each target's boxes to have shape [num_target_boxes, 4], so these samples can fail once an empty target is batched; reshape empty outputs to (0, 4) as the EoMT path does.
Useful? React with 👍 / 👎.
| } | ||
|
|
||
|
|
||
| class LTDETRInstanceSegmentationCollateFunction(_LTDETRCollateFunction): |
There was a problem hiding this comment.
This collate is never selected in the standard training flow: train_task.py instantiates collates only via train_dataset.batch_collate_fn_cls, while InstanceSegmentationDataset still points that class variable at EoMTInstanceSegmentationCollateFunction even when an LTDETR transform is installed. As a result, LTDETR instance-segmentation training through the normal dataset path silently skips the default mixup and this class's step-aware reinitialization logic.
Useful? React with 👍 / 👎.
| elif normalize == "none": | ||
| self.normalize = NormalizeArgs() |
There was a problem hiding this comment.
When users enable channel_drop with num_channels_keep other than 3 and leave normalize at the default auto, this branch still creates a 3-channel ImageNet NormalizeArgs() before num_channels is resolved from channel_drop. The sample pipeline then applies 3 mean/std values to a C!=3 image, which can fail or normalize the extra/dropped channels incorrectly; repeat/trim the stats or reject this combination as the EoMT transform compatibility pass does.
Useful? React with 👍 / 👎.
1 participant
What has changed and why?
This PR stacks on top of #835.
It adds LT-DETR instance segmentation transform support on top of the object detection transform refactor:
ltdetr_transformsinstance segmentation transform moduleHow has it been tested?
Unit tests.
Did you update CHANGELOG.md?
Did you update the documentation?