| |
| """Align OpenLane-V2 lane data timestamps/ids to nuScenes keyframes. |
| |
| OpenLane-V2 uses camera-based timestamps (~37ms offset from nuScenes lidar |
| timestamps). This script maps each lane sample to its nearest nuScenes |
| keyframe so that the lane data can be mixed into scene-sequential training |
| with det/plan/caption data sharing the same (scene_token, timestamp, id). |
| """ |
| import json |
| import sys |
| from collections import defaultdict |
| from pathlib import Path |
|
|
|
|
| def build_keyframe_lookup(det_json_path: str): |
| with open(det_json_path) as f: |
| det_data = json.load(f) |
| lookup = defaultdict(list) |
| for d in det_data: |
| lookup[d["scene_token"]].append((d["timestamp"], d["id"])) |
| for scene in lookup: |
| lookup[scene].sort() |
| return lookup |
|
|
|
|
| def find_nearest_keyframe(scene_kfs, lane_ts, max_diff_us=100_000): |
| best_diff = float("inf") |
| best_ts = None |
| best_id = None |
| for ts, sid in scene_kfs: |
| diff = abs(ts - lane_ts) |
| if diff < best_diff: |
| best_diff = diff |
| best_ts = ts |
| best_id = sid |
| if best_diff > max_diff_us: |
| return None, None, best_diff |
| return best_ts, best_id, best_diff |
|
|
|
|
| def main(): |
| det_json = sys.argv[1] if len(sys.argv) > 1 else "data/atlas_nuscenes_train.json" |
| lane_json = sys.argv[2] if len(sys.argv) > 2 else "data/openlane_subsetB_lane_train_4pt.json" |
| out_json = sys.argv[3] if len(sys.argv) > 3 else lane_json.replace(".json", "_aligned.json") |
|
|
| print(f"Det: {det_json}") |
| print(f"Lane: {lane_json}") |
| print(f"Out: {out_json}") |
|
|
| keyframe_lookup = build_keyframe_lookup(det_json) |
| with open(lane_json) as f: |
| lane_data = json.load(f) |
|
|
| aligned = [] |
| skipped_no_scene = 0 |
| skipped_too_far = 0 |
| diffs = [] |
|
|
| for lane in lane_data: |
| scene = lane["scene_token"] |
| if scene not in keyframe_lookup: |
| skipped_no_scene += 1 |
| continue |
| new_ts, new_id, diff = find_nearest_keyframe( |
| keyframe_lookup[scene], lane["timestamp"] |
| ) |
| if new_ts is None: |
| skipped_too_far += 1 |
| continue |
| lane["timestamp"] = new_ts |
| lane["id"] = new_id |
| aligned.append(lane) |
| diffs.append(diff) |
|
|
| with open(out_json, "w") as f: |
| json.dump(aligned, f, ensure_ascii=False) |
|
|
| print(f"\nResults:") |
| print(f" Aligned: {len(aligned)}") |
| print(f" Skipped (no scene): {skipped_no_scene}") |
| print(f" Skipped (too far): {skipped_too_far}") |
| if diffs: |
| print(f" Avg offset: {sum(diffs)/len(diffs):.0f} us ({sum(diffs)/len(diffs)/1e6:.4f}s)") |
| print(f" Max offset: {max(diffs)} us ({max(diffs)/1e6:.4f}s)") |
| print(f"\nSaved to: {out_json}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
