gt boxes for camera only 3d detection

xyupeng commented 2 years ago

Hi,

I want to check that the gt boxes used for the camera only 3d detection task are those appear in at least one of the camera images? This is for creating my own validation gt.bin, since I suppose the provided v1.2 gt.bin should include all lidar boxes?

xyupeng commented 2 years ago

And will you provide sth like compute_detection_metrics_main for LET-3D-AP and LET-3D-APL?

xyupeng commented 2 years ago

Hi,

May I know how the provided gt.bin is generated? I tried to generate from *.tfrecord but get a different number of objects(labels).

and is there a validation gt.bin for the camera-only 3d detection task?

Any info would be very much helpful.

VincentCa commented 2 years ago

Hi,

All LiDAR labels that are visible in at least one camera have camera_synced_box and most_visible_camera_name populated. For the 3D Camera-only Detection Challenge, at least V1.3.1 of the dataset is required, since V1.3.0 and earlier did not include all necessary metadata.
Yes, we have added compute_detection_let_metrics_main.cc to compute LET-3D-AP and LET-3D-APL from a stand-alone binary.
gt.bin is a serialized metrics.Objects proto. We did not include new versions of gt.bin in the latest releases and would recommend using the labels included in the tfrecords directly.

Hope this helps!

waymo-research / waymo-open-dataset

gt boxes for camera only 3d detection #451