README.md

---
license: apache-2.0
pretty_name: ReplicaOcc
size_categories:
- 10K<n<100K
tags:
- 3d
- robotics
- slam
- rgb-d
- occupancy
- embodied-ai
- replica
---

# Replica_OCC Benchmark

Replica_OCC is a Replica-based occupancy benchmark constructed in the data organization style of [EmbodiedOcc-ScanNet](https://huggingface.co/datasets/YkiWu/EmbodiedOcc-ScanNet) and [OccScanNet](https://huggingface.co/datasets/hongxiaoy/OccScanNet). It provides RGB-D sequences and scene-level occupancy ground truth for evaluating embodied occupancy prediction systems.

Ground-truth occupancy and poses are intended for evaluation-time alignment and metric computation. They are not required for training FreeOcc and are not used for map construction.

## Citation

If you use Replica_OCC with [FreeOcc](https://the-masses.github.io/freeocc-web/), please cite:

```bibtex
@article{jiang2026freeocc,
  title={FreeOcc: Training-Free Embodied Open-Vocabulary Occupancy Prediction},
  author={Jiang, Zeyu and Zhou, Changqing and Zuo, Xingxing and Chen, Changhao},
  journal={arXiv preprint arXiv:2604.28115},
  year={2026}
}
```

## Released Files

The released dataset package contains:

```text
Replica_OCC/
├── README.md
├── replica_name.txt
├── prepare_preprocessed.py
├── prepare_scene_occ.py
├── vis_preprocessed.py
├── vis_scene_occ.py
└── Replica_OCC/
    ├── preprocessed/
    ├── global_occ_package/
    └── sequences/
```

Only the four preparation/visualization scripts above are part of the released benchmark utilities.

## Dataset Layout

```text
Replica_OCC/
├── preprocessed/
│   ├── office0.npy
│   ├── office1.npy
│   ├── office2.npy
│   ├── office3.npy
│   ├── office4.npy
│   ├── room0.npy
│   ├── room1.npy
│   └── room2.npy
├── global_occ_package/
│   ├── office0.pkl
│   ├── office1.pkl
│   ├── office2.pkl
│   ├── office3.pkl
│   ├── office4.pkl
│   ├── room0.pkl
│   ├── room1.pkl
│   └── room2.pkl
└── sequences/
    ├── cam_params.json
    ├── office0/
    │   ├── color/
    │   │   ├── 0.jpg
    │   │   └── ...
    │   ├── depth/
    │   │   ├── 0.png
    │   │   └── ...
    │   ├── pose/
    │   │   ├── 0.txt
    │   │   └── ...
    │   └── intrinsic/
    │       ├── intrinsic_color.txt
    │       ├── intrinsic_depth.txt
    │       ├── extrinsic_color.txt
    │       └── extrinsic_depth.txt
    ├── office1/
    ├── office2/
    ├── office3/
    ├── office4/
    ├── room0/
    ├── room1/
    └── room2/
```

For FreeOcc evaluation, use:

```bash
DATA_ROOT=/path/to/Replica_OCC/sequences
SCENE_OCC_ROOT=/path/to/Replica_OCC
```

## Coordinate System

The occupancy ground truth is built in the original Replica world coordinate system.

`prepare_preprocessed.py` back-projects depth pixels using Replica camera poses and intrinsics. The resulting sparse semantic voxel points are stored in Replica world coordinates. `prepare_scene_occ.py` then builds a dense occupancy grid from those points and saves voxel center coordinates in the same Replica world coordinate system.

## Preparation Pipeline

The benchmark is created in two main stages, followed by optional visualization checks.

### 1. Build Sparse Semantic Voxels

`prepare_preprocessed.py` reads raw Replica RGB-D/semantic data and camera poses. For each scene, it back-projects depth pixels into 3D world points, assigns semantic labels from the semantic-id images, and voxelizes them by majority vote.

Input expected by the script:

```text
Replica_SLAM/
├── cam_params.json
├── office0/
│   ├── depths/depth000000.png
│   ├── semantic_ids/semantic_id000000.png
│   └── traj.txt
└── ...
```

Example command:

```bash
python prepare_preprocessed.py \
  --replica_root ./Replica_SLAM \
  --out_root ./Replica_OCC \
  --stride 4 \
  --depth_scale -1 \
  --max_depth 10.0 \
  --max_frames -1
```

Output:

```text
Replica_OCC/preprocessed/<scene>.npy
```

Each `.npy` stores an array of shape `(N, 7)`:

```text
x, y, z, r, g, b, label
```

The RGB columns are placeholders; the semantic label is stored in the last column.

### 2. Inspect Sparse Voxels

`vis_preprocessed.py` visualizes the sparse semantic voxels produced by the previous step.

```bash
python vis_preprocessed.py \
  --npy ./Replica_OCC/preprocessed/office0.npy
```

This is mainly a sanity check for depth back-projection and semantic labels.

### 3. Build Scene-Level Occupancy Packages

`prepare_scene_occ.py` converts `preprocessed/<scene>.npy` into a dense scene-level occupancy package. It builds a regular voxel grid, assigns labels by nearest-neighbor lookup, and computes an observed-space mask by projecting voxels into Replica depth frames.

Example command:

```bash
python prepare_scene_occ.py \
  --replica_root ./Replica_SLAM \
  --preprocessed_dir ./Replica_OCC/preprocessed \
  --out_dir ./Replica_OCC/global_occ_package \
  --obs_stride_frame 1 \
  --obs_stride_pix 1 \
  --mask_dilate 0 \
  --obs_max_frames -1 \
  --max_depth 10.0
```

Output:

```text
Replica_OCC/global_occ_package/<scene>.pkl
```

Each `.pkl` contains:

```text
scene_name       scene id
scene_dim        dense occupancy grid dimensions
global_pts       dense voxel centers in Replica world coordinates
global_labels    voxel labels
global_mask      observed-space mask
valid_img_count  number of depth images used for mask construction
valid_img_paths  image paths used by the mask builder
```

Label convention:

```text
0    known free space
>0   occupied semantic label
255  unknown / unobserved
```

### 4. Inspect Final Occupancy Packages

`vis_scene_occ.py` visualizes the final scene-level occupancy package.

```bash
python vis_scene_occ.py \
  --pkl ./Replica_OCC/global_occ_package/office0.pkl \
  --downsample 1
```

Visualization color meaning:

```text
mask = 0              unknown region
mask = 1, label = 0   known free space
mask = 1, label > 0   occupied semantic voxels
```

## Labels

`replica_name.txt` stores the semantic class names used by Replica_OCC. Evaluation and visualization scripts can use this file to map semantic label ids to readable names.

## Scenes

Replica_OCC contains the following eight Replica scenes:

```text
office0
office1
office2
office3
office4
room0
room1
room2
```

## Notes

- `preprocessed/*.npy` is an intermediate representation used to reproduce `global_occ_package/*.pkl`.
- `global_occ_package/*.pkl` is the occupancy ground truth used by evaluation.
- `sequences/*` is the RGB-D input used by SLAM and Gaussian mapping.
- The dataset follows a ScanNet-like RGB-D folder layout so it can be used by FreeOcc's shared dataloader.