---
pretty_name: "Mixed Signals V2X: Collaborative 3D Object Detection"
license: cc-by-nc-sa-4.0
tags:
  - 3d-object-detection
  - lidar
  - point-cloud
  - autonomous-driving
  - v2x
  - collaborative-perception
size_categories:
  - 10K<n<100K
# Data ships as .tar archives that HF cannot auto-convert to Parquet; disable
# the viewer / Data Studio so the page does not render a broken preview panel.
viewer: false
---

# Mixed Signals V2X: Collaborative 3D Object Detection Dataset

Point clouds and 3D bounding-box labels for the Mixed Signals dataset, a
diverse, real-world dataset for heterogeneous LiDAR V2X collaboration
(ICCV 2025). Collected at a busy intersection with 3 connected vehicles and
a roadside unit (RSU) carrying two LiDARs, for 5 LiDAR sensors per synchronized
frame.

This repository accompanies a collaborative 3D object detection competition on
[Codabench](https://www.codabench.org/), built on the
[Mixed Signals dataset](https://mixedsignalsdataset.cs.cornell.edu/).

## Competition

A predictions-only collaborative 3D detection challenge (Vehicle / Bike /
Pedestrian) scored by BEV rotated-box mAP. Two phases: Development (public validation split,
Aug 1 to 28 2026), then Final (held-out test split, Aug 29 to
Sep 4 2026). Download the data here, train locally, and submit per-frame
prediction files on Codabench. See the competition's starting kit for the
submission format.

## Contents

The dataset is delivered as per-segment `.tar` archives (not loose files), so
downloads stay fast and resumable. One archive is one ~30 s scene.

```
train/  mini_<N>.tar     # 29 archives: point clouds + odometry + labels
val/    mini_<N>.tar     #  4 archives: point clouds + odometry
test/   test-XXXX.tar    # ~6 pooled archives: anonymized bundles (see Splits)
```

Each train or val archive extracts to the native layout:

```
PointClouds/            # per-agent point clouds (ASCII .pcd: x y z intensity)
  mini_<N>/
    top_<k>_<ts>.pcd    dome_<k>_<ts>.pcd     # RSU lidars
    003_<k>_<ts>.pcd    004_<k>_<ts>.pcd    laser_<k>_<ts>.pcd   # vehicles
Odometry/               # per-vehicle map-frame poses (nav_msgs/Odometry CSVs)
  mini_<N>/ odometry_{003,004,laser}.csv
labels/                 # 3D box ground truth: TRAINING ARCHIVES ONLY
  mini_<N>/ mini_<N>_<k>.txt
```

Each test archive extracts to self-contained anonymized bundles, one folder
per frame, containing the 5 agent clouds plus a precomputed `transforms.json`:

```
<token>/ 003.pcd 004.pcd dome.pcd laser.pcd top.pcd  transforms.json
```

## Downloading and using

Download only the segments you need and extract in place. The archives rebuild
the folder layout the dataloader reads:

```bash
# one train segment
hf download sberrio/Mixed-Signals-V2X train/mini_6.tar --repo-type dataset --local-dir .
tar xf train/mini_6.tar          # -> PointClouds/mini_6 + Odometry/mini_6 + labels/mini_6

# the whole test split (a handful of archives)
hf download sberrio/Mixed-Signals-V2X --repo-type dataset --include "test/*" --local-dir .
for f in test/test-*.tar; do tar xf "$f" -C test/; done   # -> test/<token>/...
```

Then load frames with the starting kit's `msig_dataloader.py` (native backend for
train/val, bundle backend for test). No loader changes are needed.

## Splits

37 segments (`mini_1 … mini_37`), each a ~30 s synchronized sequence.

| Split | Segments | Labels here? |
|:--|:--|:--:|
| Train | 29 segments | ✅ yes |
| Validation | 4 segments | ❌ withheld (Development leaderboard) |
| Test | 4 segments | ❌ withheld (Final leaderboard) |

Point clouds are provided for all 37 segments. Only train labels are public.
Validation and test labels are held out for the competition leaderboards.

## Label format

`labels/mini_<N>/mini_<N>_<syncIndex>.txt`, one file per synchronized frame.
The frame id `mini_<N>_<syncIndex>` pairs with the point clouds: sync step `k`
of segment `mini_N` is the clouds `top_k…, dome_k…, 003_k…, 004_k…, laser_k…`.

Each line is one 3D box, with 8 whitespace-separated columns:

```
cls  cx  cy  cz  dx  dy  dz  yaw
```

| Field | Meaning | Units |
|:--|:--|:--|
| `cls` | class: `1`=Vehicle, `2`=Bike, `3`=Pedestrian | n/a |
| `cx cy cz` | box centre, "top" (RSU) ego frame | metres |
| `dx dy dz` | length, width, height | metres |
| `yaw` | heading (rotation about +z) | radians |

Empty files are frames with no labelled objects. Boxes are restricted to
`x, y ∈ [-51.2, 51.2] m` and contain >5 aggregated lidar points. The 10
fine-grained annotation classes are grouped into the 3 meta-classes above.

Training counts: 29 segments, 8,394 frames, 124,668 boxes
(Vehicle 48,721, Bike 30,893, Pedestrian 45,054).

## Coordinate frame

All boxes are in the "top" RSU LiDAR (ego) frame. The two RSU lidars are
static. The three vehicles' poses are in `Odometry/`. Predictions for the
competition must also be in the "top" frame.

## License and citation

Released under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/)
(attribution, non-commercial, share-alike). If you use this data, please cite:

```bibtex
@inproceedings{luo2025mixed,
  title     = {Mixed Signals: A Diverse Point Cloud Dataset for
               Heterogeneous LiDAR V2X Collaboration},
  author    = {Luo, Katie Z. and Dao, Minh-Quan and Liu, Zhenzhen and
               Campbell, Mark and Chao, Wei-Lun and Weinberger, Kilian Q. and
               Malis, Ezio and Fr\'emont, Vincent and Hariharan, Bharath and
               Shan, Mao and Worrall, Stewart and Berrio Perez, Julie Stephany},
  booktitle = {Proceedings of the IEEE/CVF International Conference on
               Computer Vision (ICCV)},
  year      = {2025}
}
```