[jacoblambert]

Summary

T4Devkit now loads every dataset through load_dataset() (#304): 1.0 datasets are read as they are, and 0.x datasets are converted to 1.0 records when loading (#307, #308). Lookups, boxes, time series and rendering all work on 1.0 records, and T4Devkit never checks the dataset version.

Breaking: the tables T4Devkit exposes are 1.0 records.

The constructor is unchanged: T4Devkit(data_root, revision=None, verbose=True), and revision still picks the version folder.

Breaking changes for users

Before (0.x records)	Now (1.0 records)
t4.calibrated_sensor	gone: calibration is on t4.sensor (translation, rotation, camera.intrinsic, camera.width / height, lidar.fields)
sample_data.calibrated_sensor_token	sample_data.sensor_token
sample_data.width / height	sensor.camera.width / height
t4.surface_ann, sample.surface_anns	rows of t4.object_ann with instance_token: None, in sample.ann_2ds
t4.lidarseg	gone; 0.x LiDAR segmentation is not loaded or rendered
instance.instance_name	instance.name
links "" (next, prev, ...)	None
render_calibration(sensor, calibration, resolution)	render_calibration(sensor)
load_table(), fix_category_table()	removed (no caller left; the category index rule is in the 0.x conversion, #308)

New tables: t4.vehicle, t4.route, t4.scenario, t4.caption.

Commits

1. feat(loader): load datasets through load_dataset()
   • Tables come from load_dataset(). get / get_table / get_idx use 1.0 table names through T4Tables (new T4Tables.get_idx), and the shortcut fields (sample.data, ann_3ds, ann_2ds, sensor.first_sd_token, ...) come from T4Tables, so T4Devkit's own reverse index is removed.
   • dataset_id / version come from manifest.json for 1.0, from the folder names for 0.x.
   • get_sample_data takes intrinsics, image size and extrinsics from the sensor row.
   • Needed to keep the samples and real data working: None links in box_velocity / get_box3ds; unknown visibility (visibility_token: None) → UNAVAILABLE; a Box2D without instance gets uuid=None.
2. feat(tier4): label confidence and time series on 1.0 records
   • The time series helper (future paths of get_box3d(future_seconds=)) walks links until None; before, every future path failed on get(None).
   • Box confidence is the label's; labels without one (manual) keep 1.0.
3. fix(sample): store a real RLE road mask in the 1.0 sample
   • The 1.0 sample's road mask (test(sample): add T4Dataset 1.0 sample dataset #295) had counts: "0", which is not a base64 RLE; the rendering test was the first to decode it. The generator now computes it with pycocotools; object_ann.json and its manifest hash are updated.
4. feat(helper): render 1.0 records and data files
   • Frames are read by their own format: LiDAR with LidarPointCloud.from_file(path, frame_index=, fields=sensor.lidar.fields) (Parquet, PCD, .bin). Camera image files are still passed as paths (sent encoded); video frames are decoded with load_image().
   • One first frame per channel, skipping sensors without frames (source LiDARs of a merged channel crashed rendering before). render_pointcloud uses the first LiDAR channel.
   • render_instance handles instances labeled only in 2D (first / last annotation is an object_ann) or with no annotation.
   • 2D masks: one loop over object_ann, former surface_ann regions included. Class IDs: category.index, or the list position when it is not set.
   • Removed: lidarseg rendering and the use of SegmentationPointCloud.

Tests

• tests/test_tier4.py: every T4Devkit test runs on both samples (0.x converted on loading, and 1.0).
• New: unknown visibility, autolabel confidence, future paths, and test_rendering (scene, instance and pointcloud of both samples rendered to .rrd files, no viewer window). There were no rendering tests before.
• tests/viewer/test_viewer.py: render_calibration with a 1.0 sensor row.

[ktro2828]

LGTM