openvinotoolkit / datumaro

Dataset Management Framework, a Python library and a CLI tool to build, analyze and manage Computer Vision datasets.
https://openvinotoolkit.github.io/datumaro/
MIT License
550 stars 135 forks source link

KITTI tracklet h, w fields swapped #1578

Open playertr opened 3 months ago

playertr commented 3 months ago

In the KITTI exporter at line 342, the h and w attributes are swapped so that they correspond the the second and first elements, respectively, of the Cuboid3D scale attribute:

track = {
                        "objectType": label,
                        "h": ann.scale[1],
                        "w": ann.scale[0],
                        "l": ann.scale[2],
                        "first_frame": frame_id,
                        "poses": [],
                        "finished": 1,  # keep last
                    }

This corresponds to the re-ordering [sy, sx, sz] with respect to the definitions from the Cuboid3D class definition.

However, this appears to be inconsistent with the definition from Geiger, 2013, p. 3, which says that length corresponds to the "x" axis of the tracklet.

Screenshot from 2024-08-01 15-49-40

Is this inconsistency expected? I needed to re-swap these dimensions when using the labels.

wonjuleee commented 2 months ago

Hi @playertr, you are correct. From the KITTI officials, they described as below. image

So we fixed this at https://github.com/openvinotoolkit/datumaro/pull/1596 and this will be published as 1.9.0 soon!