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 pytest | |
| from kitti_c2l.assets import generate_sequence_assets | |
| from kitti_c2l.generate import generate_sequence_rows | |
| from kitti_c2l.loader.dataset import ( | |
| materialize_sample, | |
| materialize_sample_from_assets, | |
| materialize_window_frame, | |
| materialize_window_frame_from_assets, | |
| ) | |
| from kitti_c2l.loader.projection import rasterize_inverse_depth | |
| from kitti_c2l.sequence_mapping import SequenceInfo | |
| from kitti_c2l.stages import STAGES | |
| from kitti_c2l.windows import generate_sequence_windows | |
| from kitti_c2l.writer import asset_rows_to_table, rows_to_table, windows_to_table | |
| def test_rasterize_inverse_depth_near_point_wins_collision(): | |
| pixels = np.array([[5.0, 5.0], [5.0, 5.0]]) | |
| depths = np.array([1000.0, 500.0]) # far, then near -- near must win the pixel | |
| depth_map = rasterize_inverse_depth(pixels, depths, image_size=(10, 10)) | |
| assert depth_map.shape == (1, 10, 10) | |
| assert depth_map[0, 5, 5] == np.float32(1.0 / 500.0) | |
| def test_rasterize_inverse_depth_clips_out_of_range(): | |
| pixels = np.array([[1.0, 1.0], [2.0, 2.0]]) | |
| depths = np.array([50.0, 9000.0]) # both outside [200, 8000] cm | |
| depth_map = rasterize_inverse_depth(pixels, depths, image_size=(10, 10)) | |
| assert np.all(depth_map == 0.0) | |
| def test_rasterize_inverse_depth_clips_out_of_bounds_pixels(): | |
| pixels = np.array([[-1.0, 5.0], [5.0, 100.0]]) | |
| depths = np.array([500.0, 500.0]) | |
| depth_map = rasterize_inverse_depth(pixels, depths, image_size=(10, 10)) | |
| assert np.all(depth_map == 0.0) | |
| def _first_row_dict(fake_kitti_root: Path) -> dict: | |
| seq_info = SequenceInfo(sequence="04", date="2011_09_30", drive="0016", start=0, end=3, frame_count=4, split="train") | |
| rows = list(generate_sequence_rows(seq_info, fake_kitti_root, stages=STAGES[:1])) | |
| table = rows_to_table(rows) | |
| return table.to_pylist()[0] | |
| def test_materialize_sample_shapes(fake_kitti_root: Path): | |
| row = _first_row_dict(fake_kitti_root) | |
| sample = materialize_sample(row, fake_kitti_root) | |
| assert sample.image.shape == (32, 32, 3) | |
| assert sample.image.dtype == np.uint8 | |
| assert sample.points.shape[1] == 3 | |
| assert sample.lidar_depth.shape == (1, 32, 32) | |
| assert sample.rotation.shape == (3, 3) | |
| assert sample.translation_cm.shape == (3,) | |
| def test_materialize_sample_ground_truth_vs_miscalibrated_differ(fake_kitti_root: Path): | |
| row = _first_row_dict(fake_kitti_root) | |
| gt_sample = materialize_sample(row, fake_kitti_root, use_ground_truth=True) | |
| miscal_sample = materialize_sample(row, fake_kitti_root, use_ground_truth=False) | |
| assert not np.allclose(gt_sample.rotation, miscal_sample.rotation) | |
| assert not np.allclose(gt_sample.translation_cm, miscal_sample.translation_cm) | |
| def _first_asset_row_dict(fake_kitti_root: Path) -> dict: | |
| seq_info = SequenceInfo(sequence="04", date="2011_09_30", drive="0016", start=0, end=3, frame_count=4, split="train") | |
| rows = list(generate_sequence_assets(seq_info, fake_kitti_root)) | |
| table = asset_rows_to_table(rows) | |
| return table.to_pylist()[0] | |
| def test_materialize_sample_from_assets_matches_from_disk(fake_kitti_root: Path): | |
| """The materialized-assets path must produce byte-identical results to reading from disk.""" | |
| row = _first_row_dict(fake_kitti_root) | |
| asset_row = _first_asset_row_dict(fake_kitti_root) | |
| assert asset_row["frame_id"] == f"{row['sequence']}_{row['frame_index']:06d}" | |
| from_disk = materialize_sample(row, fake_kitti_root) | |
| from_assets = materialize_sample_from_assets(row, asset_row) | |
| np.testing.assert_array_equal(from_disk.image, from_assets.image) | |
| np.testing.assert_array_equal(from_disk.points, from_assets.points) | |
| np.testing.assert_array_equal(from_disk.lidar_depth, from_assets.lidar_depth) | |
| np.testing.assert_array_equal(from_disk.rotation, from_assets.rotation) | |
| def _first_window_row_dict(fake_kitti_root: Path, window_length: int = 4) -> dict: | |
| seq_info = SequenceInfo(sequence="04", date="2011_09_30", drive="0016", start=0, end=3, frame_count=4, split="train") | |
| rows = list(generate_sequence_windows(seq_info, fake_kitti_root, window_length=window_length, stages=STAGES[:1])) | |
| table = windows_to_table(rows) | |
| return table.to_pylist()[0] | |
| def _asset_row_dict_for_frame(fake_kitti_root: Path, frame_index: int) -> dict: | |
| seq_info = SequenceInfo(sequence="04", date="2011_09_30", drive="0016", start=0, end=3, frame_count=4, split="train") | |
| rows = list(generate_sequence_assets(seq_info, fake_kitti_root)) | |
| table = asset_rows_to_table(rows) | |
| return next(r for r in table.to_pylist() if r["frame_index"] == frame_index) | |
| def test_materialize_window_frame_shares_decalibration_across_offsets(fake_kitti_root: Path): | |
| """Every frame_offset within a window must use the SAME (miscalibrated) extrinsic.""" | |
| window_row = _first_window_row_dict(fake_kitti_root, window_length=4) | |
| sample_0 = materialize_window_frame(window_row, frame_offset=0, kitti_raw_root=fake_kitti_root) | |
| sample_1 = materialize_window_frame(window_row, frame_offset=1, kitti_raw_root=fake_kitti_root) | |
| np.testing.assert_array_equal(sample_0.rotation, sample_1.rotation) | |
| np.testing.assert_array_equal(sample_0.translation_cm, sample_1.translation_cm) | |
| # But the actual image/points differ -- different frames, same calibration. | |
| assert not np.array_equal(sample_0.image, sample_1.image) | |
| def test_materialize_window_frame_out_of_range_raises(fake_kitti_root: Path): | |
| window_row = _first_window_row_dict(fake_kitti_root, window_length=4) | |
| with pytest.raises(IndexError): | |
| materialize_window_frame(window_row, frame_offset=4, kitti_raw_root=fake_kitti_root) | |
| def test_materialize_window_frame_from_assets_matches_from_disk(fake_kitti_root: Path): | |
| window_row = _first_window_row_dict(fake_kitti_root, window_length=4) | |
| asset_row = _asset_row_dict_for_frame(fake_kitti_root, frame_index=2) | |
| from_disk = materialize_window_frame(window_row, frame_offset=2, kitti_raw_root=fake_kitti_root) | |
| from_assets = materialize_window_frame_from_assets(window_row, asset_row) | |
| np.testing.assert_array_equal(from_disk.image, from_assets.image) | |
| np.testing.assert_array_equal(from_disk.points, from_assets.points) | |
| np.testing.assert_array_equal(from_disk.rotation, from_assets.rotation) | |