Research map / Detection and segmentation / 3D detection and driving
MSU-4S - The Michigan State University Four Seasons Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MSU-4S, a multimodal autonomous driving dataset of over 100,000 camera, lidar, and radar frames recorded across four seasons and varied weather, and quantifies domain shift with 3D (PV-RCNN++) and 2D (YOLOv5) object detection.
Problem
Existing autonomous driving datasets (KITTI, nuScenes, Waymo, BDD100K) lack colder-weather and seasonal coverage, and few offer same-location multi-season, multi-weather data with 3D labels and radar, limiting analysis of domain shift in perception.
Approach
Record the same physical routes across four seasons and weather conditions with synchronized cameras, two lidars, six radars, GNSS, and CAN, then label 2D and 3D boxes so domain shift can be measured with scene content held roughly constant.
Key results
- PV-RCNN++ trained on clear weather drops about 18 AP points when tested on snow (IoU=0.7 AP falls from 54.22 to 35.77).
- YOLOv5 2D mAP drops by 13.3 points when trained on spring rain and tested on fall clear and 19.3 points when trained on winter snow and tested on spring rain.
- Training on snow data costs only about 3-4 AP points when tested on clear weather, suggesting models trained on harder conditions transfer more robustly.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on clear-weather data and test on rain or snow instead of clear weather | PV-RCNN++ · MSU-4S 3D detection | AP (IoU=0.5; IoU=0.7): 67.99 (54.22) -> -5.55 (-10.72) on rain; 67.99 (54.22) -> -12.49 (-18.45) on snow | Hurt |
| Train on rain data and test on clear or snowy data instead of rain | PV-RCNN++ · MSU-4S 3D detection | AP (IoU=0.5; IoU=0.7): 70.09 (50.55) -> -6.50 (-3.93) on clear; 70.09 (50.55) -> -7.29 (-5.16) on snow | Hurt |
| Train on snow data and test on clear or rain instead of snow | PV-RCNN++ · MSU-4S 3D detection | AP (IoU=0.5; IoU=0.7): 68.15 (54.22) -> -3.86 (-3.24) on clear; 68.15 (54.22) -> -0.21 (-0.92) on rain | Hurt |
| Train on Spring Rain and test on Fall Clear rather than Spring Rain | YOLOv5 · MSU-4S 2D detection | mAP: 40.2; -13.30 | Hurt |
| Train on Spring Rain and test on Summer Clear rather than Spring Rain | YOLOv5 · MSU-4S 2D detection | mAP: 40.2; -11.10 | Hurt |
| Train on Spring Rain and test on Winter Snow rather than Spring Rain | YOLOv5 · MSU-4S 2D detection | mAP: 40.2; -10.40 | Hurt |
| Train on Fall Clear and test on Spring Rain rather than Fall Clear | YOLOv5 · MSU-4S 2D detection | mAP: 39.4; -6.20 | Hurt |
| Train on Fall Clear and test on Summer Clear rather than Fall Clear | YOLOv5 · MSU-4S 2D detection | mAP: 39.4; -7.30 | Hurt |
| Train on Fall Clear and test on Winter Snow rather than Fall Clear | YOLOv5 · MSU-4S 2D detection | mAP: 39.4; -6.80 | Hurt |
| Train on Summer Clear and test on Spring Rain rather than Summer Clear | YOLOv5 · MSU-4S 2D detection | mAP: 40.3; -5.60 | Hurt |
| Train on Summer Clear and test on Fall Clear rather than Summer Clear | YOLOv5 · MSU-4S 2D detection | mAP: 40.3; -9.20 | Hurt |
| Train on Summer Clear and test on Winter Snow rather than Summer Clear | YOLOv5 · MSU-4S 2D detection | mAP: 40.3; -13.40 | Hurt |
Surprising findings
- A clear-to-fall/summer domain shift is significant even though summer and fall share similar weather and lighting conditions.
- Models trained on challenging weather (snow) transfer more robustly to clear conditions than the reverse.
Limitations
- Labeling is still ongoing; object counts are estimates from a pretrained detector.
- Domain shift experiments cover only a subset of labeled segments and a single 3D detector and 2D detector.
Experimental setup
Datasets: MSU-4S
Benchmarks: MSU-4S
Metrics: AP at IoU=0.5 AP at IoU=0.7 mAP
Baselines: KITTI nuScenes Waymo BDD100K Ithica365
Builds on: PV-RCNN++ YOLOv5 ROS
Closest papers by meaning
- V2X-Radar: A Multi-modal Dataset with 4D Radar for Cooperative Perception (NeurIPS 2025 · similarity 0.83)
- L4DR: LiDAR-4DRadar Fusion for Weather-Robust 3D Object Detection (AAAI 2025 · similarity 0.82)
- USCILab3D: A Large-scale, Long-term, Semantically Annotated Outdoor Dataset (NeurIPS 2024 · similarity 0.82)
- MCD: Diverse Large-Scale Multi-Campus Dataset for Robot Perception (CVPR 2024 · similarity 0.82)
- Towards Robust 3D Object Detection with LiDAR and 4D Radar Fusion in Various Weather Conditions (CVPR 2024 · similarity 0.82)
- MAN TruckScenes: A multimodal dataset for autonomous trucking in diverse conditions (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.