Master obb

[LilBabines]

I propose an adaptation of OCSort to handle Oriented Bounding Boxes (OBB).
The bounding box (Bbox) format for the Kalman Filter is: c_x, c_y, w, h, angle.
All modifications are controlled by the OcSort.is_obb attribute. To enable OBB functionality, simply pass is_obb=True during the initialization of the OcSort tracker.

Main Modifications:

1. KalmanBoxTrackerOBB Class

   • A new class KalmanBoxTrackerOBB is created based on the KalmanBoxTracker.
   • This class uses a new Kalman Filter implementation, KalmanFilterXYWHA, which extends KalmanFilterXYSR.

2. BaseTracker Updates

   • Added an is_obb attribute to handle OBB-specific logic.
   • Updated asso_func_name :

     self.asso_func_name = asso_func + "_obb" if is_obb else asso_func

   • Adapted the BaseTracker.check_inputs() method to validate OBB inputs. OBB has 7 dimensions: c_x, c_y, w, h, angle, conf, cls (compared to 6 for standard bounding boxes).

3. OcSORT Updates

   • Updated OcSORT.update() to handle the additional OBB dimension:
     • Adjusted indices dynamically: dets[:, idx + self.is_obb].
   • Dynamically chose between KalmanBoxTrackerOBB and KalmanBoxTracker based on self.is_obb.
   • Overwrote the plotting function to handle OBB visualization.

4. Association Functions

   • Added two functions in AssociationFunction:
     • iou_batch_obb
     • centroid_batch_obb
   • Note: The iou_batch_obb function relies on an external helper function iou_obb_pair. This implementation could be improved for better integration.

5. Other Changes

   • Updated k_previous_obs() to include a new is_obb argument.
   • Added a speed_direction_obb() function used by KalmanBoxTrackerOBB.

[mikel-brostrom]

Great job here @LilBabines! I would like to add some tests before merging this first version. Do you know if a obb tracking dataset exists? I think KITTI could be projected to the ground plane but that will require quite a lot of work. Would you be okay if I used the data you provided for the tests?

[mikel-brostrom]

I believe it does make sense to check wether the len of each prediction is 7 and activate the obb tracker version based on this?

[LilBabines]

Hey @mikel-brostrom ! I don’t know of or use any public OBB dataset. I can try to find one to add a case study. I'll send a data sample once the upload is complete so you can test whatever you want for the OBB tracker.

The OBB-tracker version and the Tracker version are quite similar, which is why I tried to create an adaptive class. However, I think it makes sense to separate them, as it will also help clarify the code.

[LilBabines]

FileSender, images + labels predictions.

[mikel-brostrom]

We could just do something similar as in here:

boxmot/tests/unit/test_trackers.py

Lines 138 to 160 in 7f5dcea

	@pytest.mark.parametrize("tracker_type", PER_CLASS_TRACKERS)
	def test_per_class_tracker_active_tracks(tracker_type):
	tracker_conf = get_tracker_config(tracker_type)
	tracker = create_tracker(
	tracker_type=tracker_type,
	tracker_config=tracker_conf,
	reid_weights=WEIGHTS / 'mobilenetv2_x1_4_dukemtmcreid.pt',
	device='cpu',
	half=False,
	per_class=True
	)
	rgb = np.random.randint(255, size=(640, 640, 3), dtype=np.uint8)
	det = np.array([[144, 212, 578, 480, 0.82, 0],
	[425, 281, 576, 472, 0.72, 65]])
	embs = np.random.random(size=(2, 512))
	tracker.update(det, rgb, embs)
	# Check that tracks are created under the class tracks
	assert tracker.per_class_active_tracks[0], "No active tracks for class 0"
	assert tracker.per_class_active_tracks[65], "No active tracks for class 65"

but for obb. Let me know if you would be interested in adding this as well 😄 under tests/unit/test_trackers.py. That would complete this initial implementation.

[LilBabines]

Yes, of course! I can perform some tests for obb and add them under tests/unit/test_trackers.py.

[mikel-brostrom]

Nice 😄