Samples from BDD10K Val

Here are sample results from a pretrained YoloV3 network. Performance without tuning is surprisingly good, with a few directions for improvement:

• bounding boxes are systematically a bit too low—especially on cars and trucks, sometimes on traffic lights
• false positives of cars along the bottom of the image (frequently the dashboard of the car holding the camera is visible there)
• maximal suppression doesn't always work (nested traffic light detections)

Next steps

• display ground truth boxes
• check SOTA approaches from this paper

Fun false positive

This taxi appears to be a low-confidence bus—perhaps because of the coloration?