feat(FRNet): support t4dataset

[amadeuszsz]

Summary

Add T4dataset support for FRNet. Now you can train and run inference on T4dataset with proper multi-sensor LiDAR handling and class mapping.

Change point

• add T4SegDataset and transforms for loading T4dataset point clouds and semantic masks for single LiDAR
• add T4dataset config and deploy config
• reorganize configs and use variables for LiDAR parameters
• update README with T4dataset instructions

Note

Requires regenerating info files with create_data_t4dataset.py to include lidar_sources field.

Test performed

• Log

+---------+------------------+--------------------+----------+---------+------------+--------+--------+-------------------+--------+--------+--------------+--------------+----------------------+----------+--------+------------+---------+------------+-------------------+-----
---+-------------------+--------------+----------+--------+-------------+-------------------+--------+--------+---------+                                                                                                                                                          
| classes | drivable_surface | other_flat_surface | sidewalk | manmade | vegetation | car    | bus    | emergency_vehicle | train  | truck  | tractor_unit | semi_trailer | construction_vehicle | forklift | kart   | motorcycle | bicycle | pedestrian | personal_mobility | animal | pushable_pullable | traffic_cone | stroller | debris | other_stuff | noise+ghost_point | miou   | acc    | acc_cls |                                                                                                                                                          
+---------+------------------+--------------------+----------+---------+------------+--------+--------+-------------------+--------+--------+--------------+--------------+----------------------+----------+--------+------------+---------+------------+-------------------+--------+-------------------+--------------+----------+--------+-------------+-------------------+--------+--------+---------+                                                                                                                                                          
| results | 0.9656           | 0.6477             | 0.8462   | 0.8642  | 0.8681     | 0.9175 | 0.8217 | 0.0000            | 0.0000 | 0.8467 | 0.8184       | 0.6843       | 0.6879               | nan      | 0.0000 | 0.9138     | 0.4964  | 0.6915     | 0.0000            | nan    | 0.0934            | 0.5507       | 0.0000   | nan    | 0.7809      | 0.0352            | 0.5448 | 0.9371 | 0.6428  |                                                                                                                                                          
+---------+------------------+--------------------+----------+---------+------------+--------+--------+-------------------+--------+--------+--------------+--------------+----------------------+----------+--------+------------+---------+------------+-------------------+--------+-------------------+--------------+----------+--------+-------------+-------------------+--------+--------+---------+                                                                                                                                                          
01/28 21:07:48 - mmengine - INFO - Iter(val) [47/47]    drivable_surface: 0.9656  other_flat_surface: 0.6477  sidewalk: 0.8462  manmade: 0.8642  vegetation: 0.8681  car: 0.9175  bus: 0.8217  emergency_vehicle: 0.0000  train: 0.0000  truck: 0.8467 tractor_unit: 0.8184  semi_trailer: 0.6843  construction_vehicle: 0.6879  forklift: nan  kart: 0.0000  motorcycle: 0.9138  bicycle: 0.4964  pedestrian: 0.6915  personal_mobility: 0.0000  animal: nan  pushable_pullable: 0.0934  traffic_cone: 0.5507  stroller: 0.0000  debris: nan  other_stuff: 0.7809  noise+ghost_point: 0.0352  miou: 0.5448  acc: 0.9371  acc_cls: 0.6428  data_time: 0.0362  time: 0.1404  

Note: there was no samples where nan / zeros occurs.
Artifacts + demo in TIER IV INTERNAL LINK