Extreme memory usage loading OBB dataset

[patel-zeel]

Search before asking

• I have searched the Supervision issues and found no similar bug report.

Bug

I tried loading the DOTAv1 dataset processed by Ultralytics from this URL: https://github.com/ultralytics/assets/releases/download/v0.0.0/DOTAv1.zip with sv.DetectionDataset.from_yolo but it was unsuccessful on Google colab. When I loaded a subset of 100 instances on our compute server, it consumed nearly 36 GB of memory. Free Google colab instance comes with ~12 GB memory only so it was bound to fail there.

I investigated further and found the following reason for the extreme memory usage:

1. _with_mask function is called at the following position and it sets with_masks argument to True for OBB.
   

   supervision/supervision/dataset/formats/yolo.py

   Line 175 in 77ae682

   with_masks = _with_mask(lines=lines)

2. with_masks argument is then passed to the following function with other arguments:
   

   supervision/supervision/dataset/formats/yolo.py

   Lines 76 to 81 in 77ae682

   	def yolo_annotations_to_detections(
   	lines: List[str],
   	resolution_wh: Tuple[int, int],
   	with_masks: bool,
   	is_obb: bool = False,
   	) -> Detections:

3. Due to the effect of the with_masks argument, the following lines are executed:
   

   supervision/supervision/dataset/formats/yolo.py

   Lines 120 to 121 in 77ae682

   	mask = _polygons_to_masks(polygons=polygons, resolution_wh=resolution_wh)
   	return Detections(class_id=class_id, xyxy=xyxy, data=data, mask=mask)

4. The 2D shape of the mask is the same as the (w, h) resolution of the image. If an image has N detections, the mask shape is (N, w, h), which is not space-friendly.

The size of the mask is the main reason for extreme storage usage.

Environment

Google colab:

• Supervision: 0.25.1
• Python: 3.10.12
• Ubuntu: 22.04

Minimal Reproducible Example

Colab to reproduce the results.

Additional

I have not taken a deeper look, but a naive solution could be to generate masks only when needed and then discard them for efficient memory usage.

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!