Michigan State University Four Seasons (MSU-4S) Dataset
Version 1.2.2
Date: 07.29.2026

It was discovered that the release and archived version of the 3D labels for
the 2022_rain segment were inadvertently swapped on the file mirror. This
issue has been corrected. Release notes for the original 1.2 follow:

3D labels from our label provider are complete with postprocessing, and are
now available to download separately. Each labeled object for the frame
appears in the yaml['labels'] array, with the following structure:

label
├── box3d
│   ├── dimension
│   │   ├── x
│   │   ├── y
│   │   └── z
│   ├── location
│   │   ├── x
│   │   ├── y
│   │   └── z
│   └── orientation
│       ├── x_rotation
│       ├── y_rotation
│       └── z_rotation
├── category
├── id
├── num_pts
└── obj_id

Each object will have an ID that is unique within a given frame. However, the
"id" field may be reused across different frames for different objects.
For convenience, we added an additional obj_id field with a string ID; all
objects across frames that share the same obj_id are guaranteed to be the
same object. If you are working on applications that involve tracking, we
strongly recommend using the obj_id for this purpose.

There are some objects that are visible in the camera imagery but not in the
3D labels due to an insufficient number of lidar points. Many such objects
have no points whatsoever and are thus impossible to detect using 3D
detection alone. For consistency, we have only included objects in the 3D
labels that have a minimum of 5 points.