bag100/triangulang-scannetpp-cache

---
license: cc-by-nc-sa-4.0
tags:
- 3d
- depth
- segmentation
- scannetpp
- multi-view
size_categories:
- 100K<n<1M
---

# TrianguLang ScanNet++ Preprocessed Data

Preprocessed depth maps, camera poses, pointmaps, and rasterized semantic masks for training and evaluating [TrianguLang](https://cwru-aism.github.io/triangulang/) on [ScanNet++](https://kaldir.vc.in.tum.de/scannetpp/).

**Paper:** [arXiv:2603.08096](https://arxiv.org/abs/2603.08096)
**Code:** [github.com/bryceag11/triangulang](https://github.com/bryceag11/triangulang)
**Checkpoints:** [huggingface.co/bag100/triangulang](https://huggingface.co/bag100/triangulang)

## Contents

| Archive | Size | Description |
|---------|------|-------------|
| `da3_nested_cache_1008.tar` | 345 GB | DA3-NESTED depth + poses at 1008px (319 scenes, train+val, GT frames only) |
| `da3_nested_cache_1008_val_allframes.tar` | 79 GB | DA3-NESTED depth + poses for ALL val frames at 1008px (50 scenes) |
| `pi3xvo_cache.tar.part_a{a,b,c}` | 990 GB (3 parts) | Pi3X-VO depth + pointmaps + poses at ~672x1008 (295 scenes, train+val) |
| `semantics_2d_train.tar.part_a{a,b,c}` | 932 GB (3 parts) | Per-pixel semantic masks (263 train scenes) |
| `semantics_2d_val_v2.tar.part_a{a..f}` | 240 GB (6 parts) | Per-pixel semantic masks (51 val scenes) |

## DA3-NESTED Cache

Each `.pt` file contains:

Generated with [Depth Anything V3](https://depth-anything-3.github.io/) (DA3-NESTED-GIANT-LARGE) using overlapping chunks with Sim(3) alignment for globally consistent poses.

## Pi3X-VO Cache

Each `.pt` file contains:

Generated with [Pi3](https://pi3.github.io/) (Pi3X-VO) using chunked inference with overlapping Sim(3) alignment. Resolution is ~672x1008 (resized to 680k pixel limit with dimensions divisible by 14).

### Reassembling split archives

## Usage

Place extracted directories under your ScanNet++ data root:

Then train with:

## Semantic Masks

Per-pixel instance masks rasterized from ScanNet++ 3D mesh annotations onto DSLR images. Each `.pth` file is a numpy int32 array where pixel values are object instance IDs.

These masks are derived from the [ScanNet++](https://kaldir.vc.in.tum.de/scannetpp/) dataset. Access to the underlying ScanNet++ data requires agreeing to the ScanNet++ Terms of Use.

## Requirements

You still need the original ScanNet++ dataset for RGB images, camera intrinsics, and scene metadata. Apply for access at [kaldir.vc.in.tum.de/scannetpp](https://kaldir.vc.in.tum.de/scannetpp/).

## Citation