feat(PTv3): enable t4dataset #148

amadeuszsz · 2025-12-08T01:30:19Z

Summary

T4Dataset integration for lidarseg.

Requires #140.

Note

PTv3 is not part of mmdetection3d, and therefore its implementation in AWML consists of boilerplate code.

Test performed

Log

[2025-12-23 18:48:46,302 INFO test.py line 243 3661] Test: 1512095ff0c4e38b42ef64723ca717ed [750/750]-278195 Batch 3.565 (3.502) Accuracy 0.9247 (0.6239) mIoU 0.5597 (0.5742)                                                                                                     
[2025-12-23 18:48:46,343 INFO test.py line 260 3661] Syncing ...                                                                                                                                                                                                                   
[2025-12-23 18:48:46,344 INFO test.py line 286 3661] Val result: mIoU/mAcc/allAcc 0.5742/0.6239/0.9549                                                                                                                                                                             
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_0 - drivable_surface Result: iou/accuracy 0.9627/0.9847                                                                                                                                                                 
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_1 - other_flat_surface Result: iou/accuracy 0.6352/0.7792                                                                                                                                                               
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_2 - sidewalk Result: iou/accuracy 0.8517/0.9063                                                                                                                                                                         
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_3 - manmade Result: iou/accuracy 0.9259/0.9555                                                                                                                                                                          
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_4 - vegetation Result: iou/accuracy 0.9135/0.9637                                                                                                                                                                       
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_5 - car Result: iou/accuracy 0.9481/0.9636                                                                                                                                                                              
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_6 - bus Result: iou/accuracy 0.9490/0.9842                                                                                                                                                                              
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_7 - emergency_vehicle Result: iou/accuracy 0.9703/0.9850                                                                                                                                                                
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_8 - train Result: iou/accuracy 0.0000/0.0000                                                                                                                                                                            
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_9 - truck Result: iou/accuracy 0.9466/0.9609                                                                                                                                                                            
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_10 - tractor_unit Result: iou/accuracy 0.9218/0.9742                                                                                                                                                                    
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_11 - semi_trailer Result: iou/accuracy 0.8820/0.9434                                                                                                                                                                    
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_12 - construction_vehicle Result: iou/accuracy 0.8787/0.8911                                                                                                                                                            
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_13 - forklift Result: iou/accuracy 0.0181/0.0239                                                                                                                                                                        
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_14 - kart Result: iou/accuracy 0.0000/0.0000                                                                                                                                                                            
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_15 - motorcycle Result: iou/accuracy 0.7497/0.7626                                                                                                                                                                      
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_16 - bicycle Result: iou/accuracy 0.7354/0.7882                                                                                                                                                                         
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_17 - pedestrian Result: iou/accuracy 0.8452/0.8828                                                                                                                                                                      
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_18 - personal_mobility Result: iou/accuracy 0.0000/0.0000                                                                                                                                                               
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_19 - animal Result: iou/accuracy 0.0000/0.0000                                                                                                                                                                          
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_20 - pushable_pullable Result: iou/accuracy 0.2776/0.4690                                                                                                                                                               
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_21 - traffic_cone Result: iou/accuracy 0.5583/0.6697                                                                                                                                                                    
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_22 - stroller Result: iou/accuracy 0.1143/0.1535                                                                                                                                                                        
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_23 - debris Result: iou/accuracy 0.0000/0.0000                                                                                                                                                                          
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_24 - other_stuff Result: iou/accuracy 0.6637/0.7990                                                                                                                                                                     
[2025-12-23 18:48:46,344 INFO test.py line 289 3661] Class_25 - noise+ghost_point Result: iou/accuracy 0.1818/0.3803                                                                                                                                                               
[2025-12-23 18:48:46,344 INFO test.py line 297 3661] <<<<<<<<<<<<<<<<< End Evaluation <<<<<<<<<<<<<<<<<

Signed-off-by: Amadeusz Szymko <[email protected]>

Copilot

Pull request overview

This PR integrates T4Dataset support for PTv3 point cloud segmentation, replacing the previous hardcoded class mapping with a flexible configuration-based approach. The changes enable proper handling of T4Dataset-specific label mappings and add visualization capabilities for segmentation results.

Key changes:

Implemented flexible class mapping system to replace hardcoded label mappings
Added visualization support for point cloud segmentation results
Updated dataset paths and configurations for T4Dataset integration

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
projects/PTv3/utils/visualization.py	Replaced file I/O utilities with visualization functions for point cloud segmentation
projects/PTv3/utils/misc.py	Added `invert_class_mapping` utility for reversing class mapping dictionaries
projects/PTv3/engines/test.py	Integrated visualization and class mapping inversion into test pipeline
projects/PTv3/engines/hooks/misc.py	Removed unsupported `weights_only` parameter from `load_state_dict`
projects/PTv3/engines/hooks/evaluator.py	Updated to use inverted class mapping for logging
projects/PTv3/datasets/t4dataset.py	Refactored to use configurable class mapping and updated data loading paths
projects/PTv3/configs/semseg-pt-v3m1-0-t4dataset.py	Added comprehensive T4Dataset class mappings and updated configurations
projects/PTv3/configs/semseg-pt-v3m1-0-base.py	Reorganized configuration to use centralized class mapping variables
projects/PTv3/configs/base/default_runtime.py	Added visualization-related configuration options
projects/PTv3/README.md	Updated example commands with corrected paths and visualization options

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

projects/PTv3/datasets/t4dataset.py

projects/PTv3/configs/semseg-pt-v3m1-0-base.py

projects/PTv3/configs/_base_/default_runtime.py

KSeangTan

LGTM overall, some minor comments

KSeangTan · 2025-12-25T03:01:28Z

projects/PTv3/datasets/t4dataset.py

 @DATASETS.register_module()
 class T4Dataset(DefaultDataset):
-    def __init__(self, sweeps=10, ignore_index=-1, **kwargs):
+    def __init__(self, sweeps=10, ignore_index=-1, class_mapping={}, **kwargs):


Are you also training with 10 sweeps?

Thanks, addressed in adbd4ba

KSeangTan · 2025-12-25T03:02:08Z

projects/PTv3/datasets/t4dataset.py

        assert split in ["train", "val", "test"]
        if split == "train":
-            return os.path.join(self.data_root, "info", f"t4dataset_xx1_infos_train.pkl")
+            return os.path.join(self.data_root, "info", f"t4dataset_j6gen2_lidarseg_infos_train.pkl")


I dont think we should hardcode it here, better is to configure it in the config

Thanks, addressed in 465ba61

KSeangTan · 2025-12-25T03:03:05Z

projects/PTv3/datasets/t4dataset.py

                data_list.extend(info["data_list"])
        return data_list

+    def map_segments(self, segment, lidarseg_categories):


Typing for segment and lidarseg_categories

Thanks, addressed in be80492 and 0b1f565

KSeangTan · 2025-12-25T03:03:42Z

projects/PTv3/engines/hooks/misc.py

        self.test_last = test_last

    def after_train(self):
+        from engines.test import TESTERS


Avoid to write import statements in a function

Thanks, addressed in 7379585

KSeangTan · 2025-12-25T03:07:47Z

projects/PTv3/utils/visualization.py

+
+    # Load coordinates from saved result file
+    if logger:
+        print(f"Loading coordinates from {result_path}")


If we have a logger, please use logger.info

Thanks, addressed in 0b1f565

KSeangTan · 2025-12-25T03:08:01Z

projects/PTv3/utils/visualization.py

-        logger.info(f"Save Lines to: {file_path}")
+
+
+def get_segmentation_colors(labels, result_path, class_colors, logger=None):


Thanks, addressed in be80492 and 0b1f565

KSeangTan · 2025-12-25T03:10:16Z

projects/PTv3/utils/visualization.py

-        x = x.clone().detach().cpu().numpy()
-    assert isinstance(x, np.ndarray)
-    return x
+def create_colors_from_predictions(predictions, num_points, class_colors, logger=None):


Can we set logger to mandatory? it's a bit ugly if we need to check if logger every time

Thanks, addressed in 0b1f565

Signed-off-by: Amadeusz Szymko <[email protected]>

feat(PTv3): enable t4dataset

72c6ce4

Signed-off-by: Amadeusz Szymko <[email protected]>

amadeuszsz self-assigned this Dec 8, 2025

pre-commit-ci bot and others added 5 commits December 8, 2025 01:30

ci(pre-commit): autofix

03822da

docs(PTv3): do not use shared directory for artifacts

b5a6524

Signed-off-by: Amadeusz Szymko <[email protected]>

feat(PTv3): adjust dataset classes handling

51ac20c

Signed-off-by: Amadeusz Szymko <[email protected]>

ci(pre-commit): autofix

61b54e8

Merge branch 'main' into feat/ptv3-t4dataset

638616e

amadeuszsz marked this pull request as ready for review December 24, 2025 01:35

amadeuszsz requested review from KSeangTan, SamratThapa120 and Copilot December 24, 2025 01:35

Copilot AI reviewed Dec 24, 2025

View reviewed changes

projects/PTv3/datasets/t4dataset.py Show resolved Hide resolved

projects/PTv3/configs/semseg-pt-v3m1-0-base.py Show resolved Hide resolved

projects/PTv3/configs/_base_/default_runtime.py Show resolved Hide resolved

KSeangTan requested changes Dec 25, 2025

View reviewed changes

amadeuszsz added 6 commits December 25, 2025 15:16

feat(PTv3): info path in config

465ba61

Signed-off-by: Amadeusz Szymko <[email protected]>

fix(PTv3): add typing

be80492

Signed-off-by: Amadeusz Szymko <[email protected]>

fix(PTv3): redundant import

7379585

Signed-off-by: Amadeusz Szymko <[email protected]>

feat(PTv3): use logger & typing

0b1f565

Signed-off-by: Amadeusz Szymko <[email protected]>

fix(PTv3): remove unused sweep param

adbd4ba

Signed-off-by: Amadeusz Szymko <[email protected]>

Merge branch 'main' into feat/ptv3-t4dataset

f2dc48c

amadeuszsz requested a review from KSeangTan December 25, 2025 07:26

		logger.info(f"Save Lines to: {file_path}")


		def get_segmentation_colors(labels, result_path, class_colors, logger=None):

feat(PTv3): enable t4dataset #148

Are you sure you want to change the base?

feat(PTv3): enable t4dataset #148

Conversation

amadeuszsz commented Dec 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Note

Test performed

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

KSeangTan left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

amadeuszsz commented Dec 8, 2025 •

edited

Loading