[Feature] Support visualization and browse_dataset on nuScenes Mono-3D dataset #542
Conversation
Codecov Report
@@ Coverage Diff @@
## master #542 +/- ##
==========================================
+ Coverage 50.77% 50.85% +0.07%
==========================================
Files 197 197
Lines 14923 15038 +115
Branches 2426 2445 +19
==========================================
+ Hits 7577 7647 +70
- Misses 6851 6885 +34
- Partials 495 506 +11
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
|
Some examples: |
|
Please solve the conflicts. |
|
|
||
| def mono_cam_box2vis(cam_box): | ||
| """This is a post-processing function on the bboxes from Mono-3D task. If | ||
| we want to perform projection visualization, we need to: |
There was a problem hiding this comment.
This is mainly because Nus 3D boxes have three rotation angles (yaw, pitch, roll), but we assume the later two are \pi / 2 and 0 and only predict yaw.
| @@ -56,46 +56,31 @@ def project_pts_on_img(points, | |||
| cv2.waitKey(100) | |||
|
|
|||
|
|
|||
| def project_bbox3d_on_img(bboxes3d, | |||
There was a problem hiding this comment.
I remove this function as it is not used in the codebase. Its name is very ambiguous since we already have project_camera/lidar/depth_bbox now. Also it works almost the same as draw_lidar_bbox3d_on_img.
| return results_dict | ||
|
|
||
| def show(self, results, out_dir): | ||
| @staticmethod | ||
| def _get_data(results, key): |
There was a problem hiding this comment.
In other 3D datasets, we have these functions to enable data loading from eval_pipeline. However, as NusMono inherited from 2D datasets, I have to duplicate these functions here. Kind of ugly I think.
There was a problem hiding this comment.
Can we move it out of a specific class as a single utilized function if most of them in different classes are very similar?
There was a problem hiding this comment.
Yes. Will do this in a new PR.
| help='Whether to visualize multi-modality data. If True, we will show ' | ||
| 'both 3D point clouds with 3D bounding boxes and 2D images with ' | ||
| 'projected bounding boxes.') | ||
| '--task', |
There was a problem hiding this comment.
I let user specify the task here.
Done. |
* Add mmrotate regression test * update regression test script
As
NuScenesMonoDatasetinherits from 2DCocoDataset, to be coherent with the visualization setting in MMDet3D:eval_pipelinefor its config fileNuScenesMonoDatasetto support data loading in visualization (e.g._extract_data) followingCustom3DDatasetmmdet3d/core/bbox/structures/utils.pynamedmono_cam_box2visTo support Mono-3D in browse_dataset, I let users specify the task of the dataset instead of inferring from config (which is troublesome).
Since we don't have kitti mono config in the current codebase, I didn't test the code on it. But I believe it will work well as KittiMono inherits from NusMono.