Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
[ICML 2026 Oral] Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence
| Date | Stars |
|---|---|
| 2026-07-24 | 370 |
| 2026-07-25 | 370 |
| 2026-07-28 | 372 |
| 2026-07-30 | 372 |
| 2026-08-07 | 373 |
| 2026-08-15 | 375 |
| 2026-08-18 | 375 |
| 2026-08-20 | 376 |
| 2026-08-25 | 377 |
| 2026-08-26 | 378 |
| 2026-08-28 | 379 |
| 2026-09-01 | 381 |
| 2026-09-05 | 382 |
| 2026-09-12 | 383 |
| 2026-09-15 | 385 |
| 2026-09-19 | 385 |
| 2026-09-20 | 385 |
Today
— stars today
This week
+2 stars this week
This month
+9 stars this month
Momentum
0.0
growth rate 0.52%/day
# [ICML 2026 Oral] Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence <p align="center"> <img src="assets/teaser.png" width="100%"> </p> <p align="center"> <a href="https://arxiv.org/abs/2603.07660"><img src="https://img.shields.io/badge/arXiv-2603.07660-b31b1b.svg" alt="arXiv"></a> <a href="https://visionary-laboratory.github.io/holi-spatial/"><img src="https://img.shields.io/badge/Website-Project-blue" alt="Website"></a> <a href="https://huggingface.co/Holi-Spatial"><img src="https://img.shields.io/badge/Dataset-Holi--Spatial-green" alt="Dataset"></a> <a href="https://huggingface.co/Holi-Spatial/HoliSpatial-2M-QA-Qwen3-VL-8B"><img src="https://img.shields.io/badge/Model-Holi--Spatial-orange" alt="Model"></a> <a href="https://huggingface.co/papers/2603.07660"><img src="https://img.shields.io/badge/HuggingFace-Paper-yellow" alt="HuggingFace Paper"></a> </p> ## Todo List - [x] We release the project page of Holi-Spatial, check on [here](https://visionary-laboratory.github.io/holi-spatial/). - [x] We have released a subset of the [Holi-Spatial dataset](https://huggingface.co/Holi-Spatial), featuring **2,000+ high-quality Gaussian models**, each accompanied by its corresponding **mesh** and **scene-level 3D bounding box annotations**, together with **HoliSpatial-QA-2M**, a **2-million-scale QA dataset** built upon part of these richly annotated 3D assets. - [x] We release all model checkpoints. - [x] We will release the pipeline code in July. # Holi-Spatial Pipeline This repository contains the data curation pipeline used to turn ScanNet, ScanNet++, or DL3DV scenes into 3DGS geometry, mesh-guided masks, object/region 3D annotations, spatial QA, and LLaMA-Factory training data. The pipeline follows the three stages described in the Holi-Spatial paper: 1. Geometric optimization: DA3 depth/point cloud initialization, then PGSR/3DGS training. 2. Image-level perception: VLM class or region discovery, then SAM3 mask generation. 3. Scene-level lift and refinement: lift masks into 3D, merge/filter boxes, caption instances, and synthesize spatial QA. Paper: [https://arxiv.org/abs/2603.07660](https://arxiv.org/abs/2603.07660) ## Repository Layout - `inference_da3_scannetppv2.py`, `run_da3.sh`: unified DA3 preprocessing for ScanNet, ScanNet++, and DL3DV. - `PGSR/`: PGSR training, rendering, and mesh-to-mask code used by this pipeline. - `3dgs_train.sh`: batch PGSR/3DGS training entrypoint for ScanNet, ScanNet++, and DL3DV. - `mesh.sh`: render meshes and generate mesh-guided masks for ScanNet v2, ScanNet++, and DL3DV. - `classic_vllm.py`, `classic_region.py`: VLM-based object and functional-region label discovery. - `sam3.py`: SAM3 text-prompted mask generation. - `3d_bounding_instance_gs_rerun_da3.py`: object 2D-to-3D lifting/refinement. - `3d_bounding_instance_gs_region.py`: functional-region 2D-to-3D lifting/refinement. - `postprocess_3d_bbox_aabb.py`: floor-aligned AABB-style box post-processing before QA generation. - `scannetppv2_new.sh`, `scannetppv2_region.sh`, `scannetv2_all.sh`, `scannetv2_region.sh`, `DL3DV_new.sh`, `dl3dv_new_2k.sh`: dataset-specific annotation batch scripts. - `qa_generation/`: instance captions, spatial QA generation, QA filtering, and LLaMA-Factory conversion. ## Setup Create a CUDA-enabled Python environment, then install the common dependencies: ```bash pip install -r requirements.txt pip install -e PGSR/submodules/diff-plane-rasterization pip install -e PGSR/submodules/simple-knn ``` Install or expose the external model packages required by the pipeline: - Depth Anything 3: provides `depth_anything_3.api.DepthAnything3`. - SAM3 checkpoint/assets: `sam3.py` downloads the default checkpoint and BPE asset from `facebook/sam3` when local paths are not provided. - A VLM served through an OpenAI-compatible vLLM endpoint. All scripts default to `http://localhost:8000/v1` or `http://localhost:8000/v1/chat/completions`. ## Expected Data Layout ScanNet v2: ```t
Excerpt of 11,119 characters
Read on GitHubWould you bet a product on this? Bounded 0–100 and slow moving.
matched fp:52db2dfa4d3c7988, topic:vlm
matched fp:52db2dfa4d3c7988, topic:benchmark