--- license: mit task_categories: - image-to-3d tags: - autonomous-driving - occupancy-prediction - occupancy-forecasting - 3d - 3d-occupancy - flow-estimation - benchmark - multi-modal - nuScenes - Waymo - CARLA - OpenCOOD --- # **UniOcc**: A Unified Benchmark for Occupancy Forecasting and Prediction in Autonomous Driving ![License](https://img.shields.io/badge/license-MIT-blue.svg) [![arXiv](https://img.shields.io/badge/arXiv-2503.24381-.svg)](https://arxiv.org/abs/2503.24381) [![GitHub](https://img.shields.io/badge/GitHub-UniOcc-Blue.svg)](https://github.com/tasl-lab/UniOcc) [![HuggingFace](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-blue)](https://huggingface.co/datasets/tasl-lab/uniocc) [Paper](https://huggingface.co/papers/2503.24381) | [Project Page](https://uniocc.github.io/) | [Code](https://github.com/tasl-lab/UniOcc) UniOcc Overview

> Autonomous Driving researchers, have you ever been bothered by the fact that popular datasets all have their different > formats, and standardizing them is a pain? Have you ever been frustrated by the difficulty of just understanding > the file semantics? This challenge is even worse in the occupancy domain. But, **UniOcc is here to help.** **UniOcc** is a unified framework for occupancy forecasting, single-frame occupancy prediction, and occupancy flow estimation in autonomous driving. By integrating multiple real-world (nuScenes, Waymo) and synthetic (CARLA, OpenCOOD) datasets, UniOcc enables multi-domain training, seamless cross-dataset evaluation, and robust benchmarking across diverse driving environments. [Yuping Wang^1,2*](https://www.linkedin.com/in/yuping-wang-5a7178185/), [Xiangyu Huang³*](https://www.linkedin.com/in/xiangyu-huang-606089292), [Xiaokang Sun¹*](https://scholar.google.com/citations?user=2sWnAjQAAAAJ&hl=en), [Mingxuan Yan¹](https://waterhyacinthinnanhu.github.io/), [Shuo Xing⁴](https://shuoxing98.github.io/), [Zhengzhong Tu⁴](https://vztu.github.io/), [Jiachen Li¹](https://jiachenli94.github.io/) ¹University of California, Riverside; ²University of Michigan; ³University of Wisconsin-Madison; ⁴Texas A&M University --- ## Supported Tasks - **Occupancy Forecasting**: Predict future 3D occupancy grids over time given historical occupancies or camera inputs. - **Occupancy Prediction**: Generate detailed 3D occupancy grids from camera inputs. - **Flow Estimation**: Provides forward and backward voxel-level flow fields for more accurate motion modeling and object tracking. - **Multi-Domain Dataset Integration**: Supports major autonomous driving datasets (nuScenes, Waymo, CARLA, etc.) with consistent annotation and evaluation pipelines. - **Ground-Truth-Free Metrics**: Beyond standard IoU, introduces shape and dimension plausibility checks for generative or multi-modal tasks. - **Cooperative Autonomous Driving**: Enables multi-agent occupancy fusion and forecasting, leveraging viewpoint diversity from multiple vehicles. --- ## Pre-requisites We simplify our benchmark so you only need: - Python 3.9 or higher ```shell pip install torch torchvision pillow tqdm numpy open3d shapely matplotlib scikit-learn pickle ``` - Huggingface ```shell pip install "huggingface_hub[cli]" ``` You **do not** need: - nuscenes-devkit - waymo-open-dataset - tensorflow --- ## Dataset Download The UniOcc dataset is available on HuggingFace. The size of each dataset is as follows: | Dataset Name | Number of Scenes | Training Instances | Size (GB) | | :----------------------------------- | ---------------: | -----------------: | --------: | | NuScenes-via-Occ3D-2Hz-mini | 10 | 404 | 0.6 | | NuScenes-via-OpenOccupancy-2Hz-mini | ~ | ~ | 0.4 | | NuScenes-via-SurroundOcc-2Hz-mini | ~ | ~ | 0.4 | | NuScenes-via-OpenOccupancy-2Hz-val | 150 | 6,019 | 6.2 | | NuScenes-via-Occ3D-2Hz-val | ~ | ~ | 9.1 | | NuScenes-via-SurroundOcc-2Hz-val | ~ | ~ | 6.2 | | NuScenes-via-Occ3D-2Hz-train | 700 | 28,130 | 41.2 | | NuScenes-via-OpenOccupancy-2Hz-train | ~ | ~ | 28.3 | | NuScenes-via-SurroundOcc-2Hz-train | ~ | ~ | 28.1 | | Waymo-via-Occ3D-2Hz-mini | 10 | 397 | 0.84 | | Waymo-via-Occ3D-2Hz-val | 200 | 8069 | 15.4 | | Waymo-via-Occ3D-2Hz-train | 798 | 31,880 | 59.5 | | Waymo-via-Occ3D-10Hz-mini | 10 | 1,967 | 4.0 | | Waymo-via-Occ3D-10Hz-val | 200 | 39,987 | 74.4 | | Waymo-via-Occ3D-10Hz-train | 798 | 158,081 | 286.6 | | Carla-2Hz-mini | 2 | 840 | 1.0 | | Carla-2Hz-val | 4 | 2,500 | 2.9 | | Carla-2Hz-train | 11 | 8,400 | 9.3 | | Carla-10Hz-mini | 2 | 4,200 | 5.0 | | Carla-10Hz-val | 4 | 12,500 | 15.0 | | Carla-10Hz-train | 11 | 42,200 | 46.5 | | OPV2V-10Hz-val | 9 | 8035 | 23.5 | | OPV2V-10Hz-train | 43 | 18676 | 49.8 | | OPV2V-10Hz-test | 16 | 3629 | 9.6 | To download each dataset, use the following command (recommend you to download only the folders you need): ```shell huggingface-cli download tasl-lab/uniocc --include "NuScenes-via-Occ3D-2Hz-mini*" --repo-type dataset --local-dir ./datasets huggingface-cli download tasl-lab/uniocc --include "Carla-2Hz-train*" --repo-type dataset --local-dir ./datasets ... ``` --- ## Contents Inside each dataset, you will find the following files: ``` datasets ├── NuScenes-via-Occ3D-2Hz-mini │ ├── scene_infos.pkl │ ├── scene_001 <-- Scene Name │ │ ├── 1.npz <-- Time Step │ │ ├── 2.npz │ │ ├── ... │ ├── scene_002 │ ... ├── OpenCOOD-via-OpV2V-10Hz-val │ ├── scene_infos.pkl │ ├── scene_001 <-- Scene Name │ │ ├── 1061 <-- CAV ID │ │ │ │ ├── 1.npz <-- Time Step │ │ │ │ ├── 2.npz │ │ │ │ ├── ... │ │ │ ├── scene_002 │ ... ``` - `scene_infos.pkl`: A list of dictionaries, each containing the scene name, start and end frame, and other metadata. - `scene_XXX`: A directory containing the data for a single scenario. - `YYY.npz`: A NumPy file containing the following data for a single time step. - `occ_label`: A 3D occupancy grid (L x W x H) with semantic labels. - `occ_mask_camera`: A 3D grid (L x W x H) with binary values with `1` indicating the voxel is in the camera FOV and `0` otherwise. - `occ_flow_forward`: A 3D flow field (L x W x H x 3) with voxel flow vectors pointing to each voxel's next frame coordinate. In the last frame, flow is 0. The unit of the flow is num_voxels. - `occ_flow_backward`: A 3D flow field (L x W x H x 3) with voxel flow vectors pointing to each voxel's previous frame coordinate. In the first frame, flow is 0. The unit of the flow is num_voxels. - `ego_to_world_transformation`: A 4x4 transformation matrix from the ego vehicle to the world coordinate system. - `cameras`: A list of camera objects with intrinsic and extrinsic parameters. - `name`: The camera name (i.e. CAM_FRONT in nuScenes). - `filename`: The **relative path** to the camera image from the original datasource (i.e. nuScenes). - `intrinsics`: A 3x3 intrinsic matrix. - `extrinsics`: A 4x4 extrinsic matrix from the camera to the ego vehicle's LiDAR. - `annotations`: A list of objects with bounding boxes and class labels. - `token`: The object token, consistent with their original datasource. - `agent_to_ego`: A 4x4 transformation matrix from the object to the ego vehicle. - `agent_to_world`: A 4x4 transformation matrix from the object to the world coordinate system. - `size`: The size of the agent's bounding box in meters. (Length, Width, Height) - `category_id`: The object category (i.e. `1` for car, `4` for pedestrian, etc.) Voxel Flow Illustration

> Note: we provide the flow annotation to both dynamic voxels (agents) and static voxels (environments) in the scene. --- ## Visualizing the Dataset You can visualize the dataset using the provided `viz.py` script. For example: ```shell python uniocc_viz.py --file_path datasets/NuScenes-via-Occ3D-2Hz-mini/scene-0061/0.npz ``` In this script, we also provide the API to visualize any 3D occupancy grid, with or without a flow field. --- ## Usage ### Without Camera Images If you only need the occupancy data, you can use the provided `uniocc_dataset.py` script to load the dataset. ```python from uniocc_dataset import UniOcc dataset_carla_mini = UniOcc( data_root="datasets/Carla-2Hz-mini", obs_len=8, fut_len=12 ) dataset_nusc_mini = UniOcc( data_root="datasets/NuScenes-via-Occ3D-2Hz-mini", obs_len=8, fut_len=12 ) dataset = torch.utils.data.ConcatDataset([dataset_carla_mini, dataset_nusc_mini]) ``` ### With Camera Images If you want to use the camera images from nuScenes, Waymo or OpV2V, it is necessary to download them from the original dataset. - [nuScenes](https://www.nuscenes.org/download) - [Waymo Open Dataset v1](https://waymo.com/open/download) - Convert to KITTI format using [this tool](https://github.com/caizhongang/waymo_kitti_converter) - [OpV2V](https://ucla.app.box.com/v/UCLA-MobilityLab-OPV2V) You can then provide the root directory to the dataloader to load the camera images. ```python from uniocc_dataset import UniOcc dataset_carla_mini = UniOcc( data_root="datasets/Carla-2Hz-mini", obs_len=8, fut_len=12, datasource_root="datasets/Carla-2Hz-mini" ) dataset_nusc_mini = UniOcc( data_root="datasets/NuScenes-via-Occ3D-2Hz-mini", obs_len=8, fut_len=12, datasource_root="" # e.g. /sweeps/CAM_FRONT ) dataset_waymo_mini = UniOcc( data_root="datasets/Waymo-via-Occ3D-2Hz-mini", obs_len=8, fut_len=12, datasource_root="" # e.g. /training/image_0 ) dataset = torch.utils.data.ConcatDataset([dataset_carla_mini, dataset_nusc_mini, dataset_waymo_mini]) ``` --- ## 🚘 Occupancy Space Localization, Segmentation, Voxel Alignment, Tracking In `uniocc_utils.py`, we provide a set of utility functions for occupancy space localization, segmentation, voxel alignment, and tracking. These functions are designed to work with the voxelized occupancy grids and can be used for various tasks such as object tracking, segmentation, and motion estimation. | **Function** | **Description** | | :------------------------------------------ | :------------------------------------------------------------------------------------------------------------------------------------------- | | `GetVoxelCoordinates` | Compute the voxel indices in a 3D occupancy grid that are occupied by a transformed bounding box. | | `VoxelToCorners` | Converts voxel indices to 3D bounding-box corner coordinates for spatial visualization and geometry computations. | | `OccFrameToEgoFrame` / `EgoFrameToOccFrame` | Transforms voxel coordinates between occupancy grid space and the ego-centric coordinate frame using voxel resolution and ego center offset. | | `AlignToCentroid` | Recenters voxel coordinates by subtracting their centroid, aligning the shape around the origin. | | `RasterizeCoordsToGrid` | Converts a list of voxel coordinates into a binary 3D occupancy grid of specified dimensions. | | `Compute3DBBoxIoU` | Approximates 3D IoU by computing overlap between 2D rotated bounding boxes and comparing height extents. | | `AlignWithPCA` | Rotates voxel point clouds to align with principal axes using PCA; supports alignment with a reference PCA basis. | | `ComputeGridIoU` | Calculates voxel-wise binary IoU between two occupancy grids of identical shape. | | `SegmentVoxels` | Performs 3D connected-component labeling (CCL) on an occupancy grid, with filtering by minimum voxel count. | | `EstimateEgoMotionFromFlows` | Estimates ego-motion from voxel flow fields over time by extracting static voxels and applying RANSAC-based rigid transform fitting. | | `AccumulateTransformations` | Composes a sequence of frame-to-frame transformation matrices into global poses over time. | | `TrackOccObjects` | Tracks objects across frames using voxel flows and estimated ego-motion, returning per-object trajectories and voxel groupings. | | `BipartiteMatch` | Solves the optimal assignment problem (Hungarian algorithm) to associate predicted and reference objects based on a cost/score matrix. | ## 👀 Visualization API In `uniocc_viz.py`, we provide a set of visualization functions to render occupancy grids, flow fields, and camera images in 3D using Open3D. These functions can be used to visualize occupancy grids, flow vectors, and the ego vehicle model in a 3D scene. | **Function** | **Description** | | :-------------------------- | :---------------------------------------------------------------------------------------------------------- | | `__voxel_to_points__` | Converts a 3D voxel array with a boolean occupancy mask into 3D point cloud coordinates, their values, and indices. | | `__voxel_profile__` | Creates bounding box profiles for each voxel as `[x, y, z, w, l, h, yaw]` for rendering. | | `__rotz__` | Computes a Z-axis rotation matrix given an angle in radians. | | `__compute_box_3d__` | Calculates 8-corner 3D box coordinates from voxel centers, dimensions, and yaw angles. | | `__generate_ego_car__` | Produces a voxelized point cloud representation of the ego vehicle centered at the origin. | | `__place_ego_car_at_position__` | Translates ego vehicle voxels to a specified center location in the scene. | | `FillRoadInOcc` | Ensures the bottom-most slice of the occupancy grid contains labeled road voxels for consistent visualization. | | `CreateOccHandle` | Builds a full Open3D visualizer, rendering occupancy grids with color and optional bounding boxes. | | `AddFlowToVisHandle` | Draws flow vectors on the Open3D visualizer as red line segments for motion inspection. | | `AddCenterEgoToVisHandle` | Adds the ego vehicle model to the visualizer at the center of the occupancy scene. | | `VisualizeOcc` | Creates a visualizer for a static occupancy grid and optionally adds the ego vehicle. | | `VisualizeOccFlow` | Visualizes both occupancy and voxel-level flow vectors in 3D, with optional ego car rendering. | | `VisualizeOccFlowFile` | Loads `.npz` files containing occupancy and flow data, and visualizes them together. | | `RotateO3DCamera` | Loads Open3D camera parameters from a JSON file and applies them to the current visualizer view. | ## 🌀 Voxel Flow Computation `uniocc_flow_gen.py` provides utility functions for computing voxel-level flow fields from occupancy grids and object annotations. | Function | Description | | :--------------------------- | :------------------------------------------------------------------------------------------------------------------------------------- | | `ComputeFlowsForObjects` | Computes flow vectors for all foreground (dynamic) objects by transforming their voxels from current to next frame using SE(3) poses. | | `ComputeFlowsForBackground` | Computes static scene flow due to ego-motion for voxels labeled as background (e.g., road, terrain). | | `ComputeFlowsForOccupancyGrid` | Combines dynamic and static voxel-level flow to produce a complete scene flow grid. | Run the script directly to compute and verify flow fields on sample data from the UniOcc dataset: ```shell python uniocc_flow_gen.py ``` ## 📏 Evaluation Demo (needs a sample dataset in `datasets/`): ```shell python uniocc_eval.py ``` We provide these evaluation APIs, as described in our paper. | Function | Description | | :---------- | :---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `FindGMMForCategory` | Fits the best Gaussian-Mixture Model (GMM) to all length-width-height triples of a chosen class, creating a “realism” prior for that object category. | | `ComputeObjectLikelihoods` | Segments each object in a binary occupancy grid and scores its bounding-box dimensions against the pretrained GMM, returning plausibility probabilities and counts. | | `ComputeTemporalShapeConsistencyByTracking` | Tracks every object across frames using voxel flows, aligns shapes, and reports the mean IoU of consecutive shapes, higher = smoother temporal geometry. | | `ComputeStaticConsistency` | Warps static voxels from frame *t* to *t+1* via ego motion and measures how well they overlap, giving an IoU-style score for background stability. | | `ComputeIoU` | Computes the standard intersection-over-union between two mono-label occupancy grids while ignoring a specified “free-space” label. | | `ComputeIoUForCategory` | Same as `ComputeIoU`, but restricted to voxels of a single semantic class, enabling per-category performance evaluation. | --- ## Checklist - [x] Release non-cooperative datasets - [x] Release cooperative dataset - [x] Release the dataset API - [x] Release the visualization script - [x] Release the evaluation scripts - [x] Release the occupancy segmentation, localization, tracking scripts - [x] Release data generation scripts --- ## Change History - **2025-07-17**: Fix bug in nuScenes datasets where `token` is incorrectly set to annotation token but should be instance token. We changed the `token` to the instance token in the `annotations` field of the occupancy grid. This change is backward compatible, as we also provide the original annotation token in the `annotation_token` field. ## Citation If you find this work useful, please consider citing our paper: ```bibtex @inproceedings{wang2025uniocc, title={UniOcc: A Unified Benchmark for Occupancy Forecasting and Prediction in Autonomous Driving}, author={Wang, Yuping and Huang, Xiangyu and Sun, Xiaokang and Yan, Mingxuan and Xing, Shuo and Tu, Zhengzhong and Li, Jiachen}, booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)}, year={2025}, publisher={IEEE} } ```