Release KITTI-C2L v1.0.0: documented self-contained benchmark, loader and baselines
db613b6 verified | from __future__ import annotations | |
| from pathlib import Path | |
| import numpy as np | |
| import pyarrow.parquet as pq | |
| from kitti_c2l.assets import generate_frame_asset, generate_sequence_assets | |
| from kitti_c2l.kitti_calib import load_rig_chain | |
| from kitti_c2l.schema import ASSETS_COLUMN_NAMES, ASSETS_SCHEMA, frame_id | |
| from kitti_c2l.sequence_mapping import SequenceInfo | |
| from kitti_c2l.writer import write_assets, write_assets_sharded | |
| def _fake_seq_info(frame_count: int = 4) -> SequenceInfo: | |
| return SequenceInfo(sequence="04", date="2011_09_30", drive="0016", start=0, end=frame_count - 1, frame_count=frame_count, split="train") | |
| def test_frame_id_format(): | |
| assert frame_id("04", 2) == "04_000002" | |
| def test_generate_frame_asset_matches_raw_files(fake_kitti_root: Path): | |
| seq_info = _fake_seq_info() | |
| rig = load_rig_chain(fake_kitti_root / "2011_09_30", camera_id=2) | |
| row = generate_frame_asset(seq_info, fake_kitti_root, frame_index=1, image_width=rig.image_width, image_height=rig.image_height) | |
| raw_image_path = fake_kitti_root / "2011_09_30" / "2011_09_30_drive_0016_sync" / "image_02" / "data" / "0000000001.png" | |
| raw_lidar_path = fake_kitti_root / "2011_09_30" / "2011_09_30_drive_0016_sync" / "velodyne_points" / "data" / "0000000001.bin" | |
| assert row.image_png == raw_image_path.read_bytes() | |
| assert row.points_bin == raw_lidar_path.read_bytes() | |
| assert row.frame_id == "04_000001" | |
| assert row.raw_frame_index == 1 | |
| def test_generate_sequence_assets_count(fake_kitti_root: Path): | |
| seq_info = _fake_seq_info() | |
| rows = list(generate_sequence_assets(seq_info, fake_kitti_root)) | |
| assert len(rows) == 4 | |
| frame_ids = {row.frame_id for row in rows} | |
| assert len(frame_ids) == 4 # one row per unique frame, no stage multiplication | |
| def test_write_assets_round_trip(fake_kitti_root: Path, tmp_path: Path): | |
| seq_info = _fake_seq_info() | |
| rows = list(generate_sequence_assets(seq_info, fake_kitti_root)) | |
| output_path = tmp_path / "frames_train.parquet" | |
| count = write_assets(rows, output_path) | |
| assert count == 4 | |
| table = pq.read_table(output_path) | |
| assert table.schema.equals(ASSETS_SCHEMA) | |
| read_rows = table.to_pylist() | |
| assert set(read_rows[0].keys()) == set(ASSETS_COLUMN_NAMES) | |
| for original, read_back in zip(rows, read_rows, strict=True): | |
| assert read_back["frame_id"] == original.frame_id | |
| assert bytes(read_back["image_png"]) == original.image_png | |
| assert bytes(read_back["points_bin"]) == original.points_bin | |
| def test_write_assets_sharded_splits_by_target_size(fake_kitti_root: Path, tmp_path: Path): | |
| seq_info = _fake_seq_info() | |
| rows = list(generate_sequence_assets(seq_info, fake_kitti_root)) | |
| row_bytes = len(rows[0].image_png) + len(rows[0].points_bin) | |
| # Force multiple shards: cap each shard at roughly 1 row's worth of bytes. | |
| shard_paths = write_assets_sharded(rows, tmp_path, split="train", target_shard_bytes=row_bytes) | |
| assert len(shard_paths) == 4 # one row per shard | |
| for i, path in enumerate(shard_paths): | |
| assert path.name == f"frames_train-{i:05d}-of-00004.parquet" | |
| assert path.exists() | |
| all_rows = [] | |
| for path in shard_paths: | |
| all_rows.extend(pq.read_table(path).to_pylist()) | |
| assert len(all_rows) == 4 | |
| assert {r["frame_id"] for r in all_rows} == {row.frame_id for row in rows} | |
| def test_write_assets_sharded_single_shard_when_under_target(fake_kitti_root: Path, tmp_path: Path): | |
| seq_info = _fake_seq_info() | |
| rows = list(generate_sequence_assets(seq_info, fake_kitti_root)) | |
| shard_paths = write_assets_sharded(rows, tmp_path, split="test", target_shard_bytes=10**9) | |
| assert len(shard_paths) == 1 | |
| assert shard_paths[0].name == "frames_test-00000-of-00001.parquet" | |
| def test_write_assets_sharded_empty_input(tmp_path: Path): | |
| shard_paths = write_assets_sharded([], tmp_path, split="train") | |
| assert shard_paths == [] | |
| def test_asset_points_bin_decodes_to_same_points_as_raw_file(fake_kitti_root: Path): | |
| seq_info = _fake_seq_info() | |
| rows = list(generate_sequence_assets(seq_info, fake_kitti_root)) | |
| row = rows[2] | |
| raw_lidar_path = fake_kitti_root / "2011_09_30" / "2011_09_30_drive_0016_sync" / "velodyne_points" / "data" / "0000000002.bin" | |
| expected = np.fromfile(raw_lidar_path, dtype=np.float32).reshape(-1, 4) | |
| decoded = np.frombuffer(row.points_bin, dtype=np.float32).reshape(-1, 4) | |
| np.testing.assert_array_equal(decoded, expected) | |