README.md

W-CODA2024 Track 2 Dataset

Dataset Description

This dataset contains auxiliary data files for the W-CODA (Multimodal Perception and Comprehension of Corner Cases in Autonomous Driving) Track 2 workshop at ECCV 2024. The files provide metadata about the nuScenes validation set for evaluating video generation and detection/segmentation results.

Data Files

1. nuscenes_infos_temporal_val_3keyframes.pkl

   • Contains information about key frames from 150 scenes in the nuScenes validation set.
   • Each scene has 3 key frames extracted from the first 16 frames.
   • Used for evaluating object detection and segmentation performance on the key frames.
   • Format: Python pickle file, load with mmcv.load().

2. nuscenes_infos_temporal_val_12hz.pkl

   • Contains metadata for 150 scenes in the nuScenes validation set.
   • Provides the first 16 frames (at 12Hz) for each scene.
   • Paths to real video frames, used to calculate Fréchet Video Distance (FVD) between generated and real videos.
   • Format: Python pickle file, load with mmcv.load().

data = mmcv.load("./data/nuscenes/nuscenes_infos_temporal_val_3keyframes.pkl")
print(data.keys())
# dict_keys(['infos', 'metadata'])

Dataset Structure

The loaded pickle files contain a list of dictionaries, one per frame, with 'scene_token', 'frame_idx', 'gt_boxes', 'gt_names', 'cams', etc. Their meanings are consistent with the original nus annotations.

Download

Download the files from Hugging Face:

• nuscenes_infos_temporal_val_3keyframes.pkl
• nuscenes_infos_temporal_val_12hz.pkl