README.md

---
license: cc-by-nc-4.0
tags:
  - 3d-object-detection
  - incremental-learning
  - point-cloud
  - scannet
  - sunrgbd
annotations_creators:
  - found
language: []
pretty_name: LDMR dataset metadata (ScanNetV2 + SUN RGB-D annotation indices)
---

# LDMR — dataset metadata

Annotation-index files (`.pkl`) for

> **Breaking the Model Forgetting Cycle in Long-Incremental 3D Object Detection**
> Peisheng Qian, Jie Xu, Xulei Yang, Na Zhao
> *European Conference on Computer Vision (ECCV), 2026*

Code: <https://github.com/qianpeisheng/LDMR> &nbsp;·&nbsp; Checkpoints: [Peisheng/LDMR](https://huggingface.co/Peisheng/LDMR)

## What this is — and what it is not

These are **annotation indices only**: per-scene bounding boxes, class ids, and
*relative* paths pointing at point-cloud files you must produce yourself. They
contain no scan data, no point clouds, and no images.

This repository does **not** redistribute ScanNet or SUN RGB-D. ScanNet requires
each user to sign its Terms of Use directly with the ScanNet team, so mirroring
the scans — or point clouds derived from them — would breach those terms.

To use these files you must download the scans from their original sources and
extract them yourself:

- **ScanNetV2** — <http://www.scan-net.org/> (requires signing the Terms of Use)
- **SUN RGB-D** — <https://rgbd.cs.princeton.edu/>

The point of publishing them is to let you skip the index-building step of data
preparation, and to pin the exact annotation set behind the paper's numbers. The
per-scene extraction step is still required.

## Contents

~37 MB.

| File | Scenes | `annos['class']` range | Notes |
|---|---|---|---|
| `scannet/scannet_infos_train_40class_corrected.pkl` | 1201 | 1–40 | class ids corrected to 1-based NYU40 |
| `scannet/scannet_infos_val_40class_corrected.pkl` | 312 | 1–40 | " |
| `scannet/scannet_infos_test_40class_corrected.pkl` | 100 | — | no annotations (test split) |
| `sunrgbd/sunrgbd_infos_train_40class.pkl` | 5285 | 0–39 | 478 scenes have `gt_num == 0` |
| `sunrgbd/sunrgbd_infos_val_40class.pkl` | 5050 | 0–39 | 436 scenes have `gt_num == 0` |

Each file unpickles to a `list[dict]`, one entry per scene, with `point_cloud`
and a relative `pts_path` (e.g. `points/scene0191_00.bin`). Except for the
ScanNet test split, each entry also has an `annos` block holding
`gt_boxes_upright_depth`, `class`, `gt_num` and related fields. ScanNet entries
additionally carry `axis_align_matrix`, `pts_instance_mask_path` and
`pts_semantic_mask_path`; SUN RGB-D entries carry `calib` and `image`.

Two details will bite you if you read these files directly.

**The two datasets use different class-id bases.** The `_corrected` suffix on the
ScanNet files is load-bearing: `tools/create_data.py` writes `annos['class']` as
0-based indices, whereas the class mappings and `valid_cat_ids` throughout the
LDMR codebase treat a ScanNet class id as the NYU40 id itself, which is 1-based.
These files have the shift applied, by
`tools/data_converter/scannet_correct_class_ids.py`; using uncorrected indices
silently mislabels every box by one class. The SUN RGB-D files are 0-based (0–39)
and need no such correction — that asymmetry is expected, and the configs handle
each dataset on its own terms.

**Object-free scenes omit the annotation arrays.** In SUN RGB-D, 478 train and
436 val scenes have `gt_num == 0`; for those the `annos` block has no `class` or
`gt_boxes_upright_depth` key at all, rather than an empty array. Guard with
`if 'class' in info['annos']` when iterating.

## Usage

```python
from huggingface_hub import snapshot_download

snapshot_download('Peisheng/LDMR-data', repo_type='dataset', local_dir='data')
```

The LDMR configs expect the files at:

```
data/scannet/scannet_infos_{train,val,test}_40class_corrected.pkl
data/sunrgbd/sunrgbd_infos_{train,val}_40class.pkl
```

alongside the point clouds you extracted from the original scans. See the
dataset-preparation section of the
[GitHub README](https://github.com/qianpeisheng/LDMR#dataset-preparation).

## License

The index files are released under
[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/), matching the
LDMR codebase. They are derived from the ScanNetV2 and SUN RGB-D annotations;
the original datasets remain governed by their own licenses and terms of use,
which you accept with their respective providers. Non-commercial use only.

## Citation

```bibtex
@inproceedings{qian2026ldmr,
  title     = {Breaking the Model Forgetting Cycle in Long-Incremental 3D Object Detection},
  author    = {Qian, Peisheng and Xu, Jie and Yang, Xulei and Zhao, Na},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
```

Please also cite ScanNet and SUN RGB-D if you use these files.