deepdetect
deepdetect copied to clipboard
feat: more RT-DETRv2 options
new options:
--num_queries to change num_queries and num_top_queries, num_denoising is set to 1/3 of this value
--img_width & --img_height to change the eval_spatial_size
--nms_threshold to add an optional NMS after inference with the specified IoU threshold
empty bboxes are now handled as well, but still fail if the whole batch contains images without bboxes maybe related: https://github.com/pytorch/pytorch/issues/144965