nvs2colmap 0.7.0__tar.gz → 0.7.3__tar.gz

{nvs2colmap-0.7.0 → nvs2colmap-0.7.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nvs2colmap
-Version: 0.7.0
+Version: 0.7.3
 Summary: Utilities for converting novel view synthesis datasets to COLMAP format.
 Author-email: Howard Yin <yindaheng98@gmail.com>
 Maintainer-email: Howard Yin <yindaheng98@gmail.com>
@@ -94,16 +94,14 @@ Install COLMAP from conda-forge:
 conda install -c conda-forge colmap
 ```
 
-The `colmap` executable is installed into the conda environment. Activate
-that environment before a `--use-colmap` run. Activation puts `colmap` on
-`PATH` and loads the shared libraries the executable links against. A Python
-virtualenv such as `.venv` can be activated afterwards so its `python` stays
-first on `PATH`.
+The `colmap` executable is `${CONDA_EXE%/bin/conda}/bin/colmap`. It loads
+shared libraries from the conda `lib` directory next to that binary, so pass
+this path to `--colmap-executable`.
 
 For Neural 3D Video scenes, pass `data/ffmpeg` and `data/ffprobe`, or put
 `ffmpeg` and `ffprobe` on `PATH`. To run the COLMAP reconstruction that
 [InstantSplat](https://github.com/yindaheng98/InstantSplat) continues from,
-pass `--colmap-executable colmap` after the conda environment is active.
+pass `--colmap-executable "${CONDA_EXE%/bin/conda}/bin/colmap"`.
 
 Extract a Neural 3D Video scene and write per-frame COLMAP text models for
 [InstantSplat](https://github.com/yindaheng98/InstantSplat):
@@ -131,7 +129,6 @@ Run COLMAP reconstruction and undistortion for each frame, for
 [InstantSplat](https://github.com/yindaheng98/InstantSplat):
 
 ```bash
-conda activate
 python -m nvs2colmap.n3dv \
   --path data/Robo360/xarm6_gold_rope_in_basket_2 \
   --ffmpeg ffmpeg \
@@ -139,7 +136,7 @@ python -m nvs2colmap.n3dv \
   --video-extension MP4 \
   --n-frames 1 \
   --use-colmap \
-  --colmap-executable colmap \
+  --colmap-executable "${CONDA_EXE%/bin/conda}/bin/colmap" \
   --colmap-use-gpu 1
 ```
 
@@ -164,7 +161,8 @@ python -m nvs2colmap.stnerf \
 
 Link a Siggraph Asia 2026 VVC scene and write per-frame COLMAP text models
 for [InstantSplat](https://github.com/yindaheng98/InstantSplat). The default
-starts at the first frame file, so `selfcap` `000900.jpg` is `frame901`:
+starts at the first frame file and writes `frame1`, including a `selfcap`
+scene whose first file is `000900.jpg`:
 
 ```bash
 python -m nvs2colmap.SiggraphAsia2026VVC \
@@ -215,8 +213,8 @@ in every extracted frame. `--use-colmap` writes images to `frame*/input` and
 masks to `frame*/input_mask`, then runs COLMAP undistortion into `images/` and
 `image_masks/` for
 [InstantSplat](https://github.com/yindaheng98/InstantSplat). The default
-starts at the first frame file. Source file `000000.jpg` is frame `1`, and
-`000900.jpg` is frame `901`.
+starts at the first frame file and writes `frame1`. A scene whose first file
+is `000900.jpg` still begins at `frame1`.
 
 For Blender scenes, rendering writes one video per camera beside
 `poses_bounds.npy`, plus `blender_frames.json` in that same directory. The

{nvs2colmap-0.7.0 → nvs2colmap-0.7.3}/README.md

@@ -65,16 +65,14 @@ Install COLMAP from conda-forge:
 conda install -c conda-forge colmap
 ```
 
-The `colmap` executable is installed into the conda environment. Activate
-that environment before a `--use-colmap` run. Activation puts `colmap` on
-`PATH` and loads the shared libraries the executable links against. A Python
-virtualenv such as `.venv` can be activated afterwards so its `python` stays
-first on `PATH`.
+The `colmap` executable is `${CONDA_EXE%/bin/conda}/bin/colmap`. It loads
+shared libraries from the conda `lib` directory next to that binary, so pass
+this path to `--colmap-executable`.
 
 For Neural 3D Video scenes, pass `data/ffmpeg` and `data/ffprobe`, or put
 `ffmpeg` and `ffprobe` on `PATH`. To run the COLMAP reconstruction that
 [InstantSplat](https://github.com/yindaheng98/InstantSplat) continues from,
-pass `--colmap-executable colmap` after the conda environment is active.
+pass `--colmap-executable "${CONDA_EXE%/bin/conda}/bin/colmap"`.
 
 Extract a Neural 3D Video scene and write per-frame COLMAP text models for
 [InstantSplat](https://github.com/yindaheng98/InstantSplat):
@@ -102,7 +100,6 @@ Run COLMAP reconstruction and undistortion for each frame, for
 [InstantSplat](https://github.com/yindaheng98/InstantSplat):
 
 ```bash
-conda activate
 python -m nvs2colmap.n3dv \
   --path data/Robo360/xarm6_gold_rope_in_basket_2 \
   --ffmpeg ffmpeg \
@@ -110,7 +107,7 @@ python -m nvs2colmap.n3dv \
   --video-extension MP4 \
   --n-frames 1 \
   --use-colmap \
-  --colmap-executable colmap \
+  --colmap-executable "${CONDA_EXE%/bin/conda}/bin/colmap" \
   --colmap-use-gpu 1
 ```
 
@@ -135,7 +132,8 @@ python -m nvs2colmap.stnerf \
 
 Link a Siggraph Asia 2026 VVC scene and write per-frame COLMAP text models
 for [InstantSplat](https://github.com/yindaheng98/InstantSplat). The default
-starts at the first frame file, so `selfcap` `000900.jpg` is `frame901`:
+starts at the first frame file and writes `frame1`, including a `selfcap`
+scene whose first file is `000900.jpg`:
 
 ```bash
 python -m nvs2colmap.SiggraphAsia2026VVC \
@@ -186,8 +184,8 @@ in every extracted frame. `--use-colmap` writes images to `frame*/input` and
 masks to `frame*/input_mask`, then runs COLMAP undistortion into `images/` and
 `image_masks/` for
 [InstantSplat](https://github.com/yindaheng98/InstantSplat). The default
-starts at the first frame file. Source file `000000.jpg` is frame `1`, and
-`000900.jpg` is frame `901`.
+starts at the first frame file and writes `frame1`. A scene whose first file
+is `000900.jpg` still begins at `frame1`.
 
 For Blender scenes, rendering writes one video per camera beside
 `poses_bounds.npy`, plus `blender_frames.json` in that same directory. The

nvs2colmap-0.7.3/nvs2colmap/SiggraphAsia2026VVC/README.md

@@ -0,0 +1,379 @@
+# Siggraph Asia 2026 Volumetric Video Challenge
+
+This package converts scenes from the **Siggraph Asia 2026 Volumetric Video
+Challenge** into per-frame COLMAP outputs. By default it writes COLMAP text
+models. Views whose distortion is zero are hardlinked. Views with a nonzero
+`D_<view>` keep a hardlink of the original in `input/`. COLMAP's
+`image_undistorter` writes `images/` and a `PINHOLE` model fitted to that
+image. Included masks follow the images. With `--use-colmap`, it runs the full
+COLMAP pipeline on every extracted frame.
+
+## Supported Datasets
+
+- **Siggraph Asia 2026 VVC**: dataset
+  [zju3dv/SiggraphAsia2026VVC](https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC).
+  The public release contains the `siga` and `selfcap` tracks. The camera YAML
+  is the EasyVolCap layout used by the previous challenge release and by
+  [TrackerSplat's SIGA2025 loader](https://github.com/yindaheng98/TrackerSplat/blob/SIGA2025VVC/tools/parse_camera_SIGA2025VVC.py).
+
+The dataset is gated and about 309 GB. Request access on the dataset page,
+then log in with `hf auth login`.
+
+Runtime dependencies:
+
+```bash
+pip install numpy torch
+pip install -U huggingface_hub
+```
+
+## Quick Start
+
+Each scene is one tar. `curl` pipes it into `tar`, so the archive is not stored.
+The archive keeps its scene directory, so the files land at
+`data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1.tar/004_1_seq1`.
+`cp -al` then hardlinks every file into
+`data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1`. Remove only that
+directory, then hardlink again. That deletes generated frames and leaves the
+extracted files in the `.tar` directory. `release_manifest.json` lists the
+public views.
+
+```bash
+mkdir -p \
+  data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1.tar \
+  data/SiggraphAsia2026VVC/siga/compression/test/006_1_seq1.tar \
+  data/SiggraphAsia2026VVC/siga/compression/test/007_0_seq1.tar \
+  data/SiggraphAsia2026VVC/siga/compression/test/008_2_seq1.tar \
+  data/SiggraphAsia2026VVC/siga/compression/test/011_0_seq1.tar \
+  data/SiggraphAsia2026VVC/siga/compression/val/001_1_seq0.tar \
+  data/SiggraphAsia2026VVC/siga/compression/val/012_0_seq0.tar \
+  data/SiggraphAsia2026VVC/siga/sparse/test/004_1_seq0.tar \
+  data/SiggraphAsia2026VVC/siga/sparse/test/006_1_seq0.tar \
+  data/SiggraphAsia2026VVC/siga/sparse/test/007_0_seq0.tar \
+  data/SiggraphAsia2026VVC/siga/sparse/test/009_0_seq0.tar \
+  data/SiggraphAsia2026VVC/siga/sparse/test/011_0_seq0.tar \
+  data/SiggraphAsia2026VVC/siga/sparse/val/001_1_seq0.tar \
+  data/SiggraphAsia2026VVC/siga/sparse/val/012_0_seq0.tar \
+  data/SiggraphAsia2026VVC/selfcap/compression/test/0512_bike.tar \
+  data/SiggraphAsia2026VVC/selfcap/compression/test/0525_corgi.tar \
+  data/SiggraphAsia2026VVC/selfcap/compression/test/0811_yoga.tar \
+  data/SiggraphAsia2026VVC/selfcap/sparse/test/0512_bike.tar \
+  data/SiggraphAsia2026VVC/selfcap/sparse/test/0525_corgi.tar \
+  data/SiggraphAsia2026VVC/selfcap/sparse/test/0811_yoga.tar
+hf auth login
+hf download zju3dv/SiggraphAsia2026VVC --type dataset --local-dir data/SiggraphAsia2026VVC --include "license.txt" --include "checksums.sha256" --include "release_manifest.json" --include "archive_manifest.json" --include "heldout_test_cameras/*"
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/compression/test/004_1_seq1.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/compression/test/006_1_seq1.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/compression/test/006_1_seq1.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/compression/test/007_0_seq1.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/compression/test/007_0_seq1.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/compression/test/008_2_seq1.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/compression/test/008_2_seq1.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/compression/test/011_0_seq1.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/compression/test/011_0_seq1.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/compression/val/001_1_seq0.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/compression/val/001_1_seq0.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/compression/val/012_0_seq0.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/compression/val/012_0_seq0.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/sparse/test/004_1_seq0.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/sparse/test/004_1_seq0.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/sparse/test/006_1_seq0.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/sparse/test/006_1_seq0.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/sparse/test/007_0_seq0.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/sparse/test/007_0_seq0.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/sparse/test/009_0_seq0.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/sparse/test/009_0_seq0.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/sparse/test/011_0_seq0.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/sparse/test/011_0_seq0.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/sparse/val/001_1_seq0.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/sparse/val/001_1_seq0.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/siga/sparse/val/012_0_seq0.tar | tar -xf - -C data/SiggraphAsia2026VVC/siga/sparse/val/012_0_seq0.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/selfcap/compression/test/0512_bike.tar | tar -xf - -C data/SiggraphAsia2026VVC/selfcap/compression/test/0512_bike.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/selfcap/compression/test/0525_corgi.tar | tar -xf - -C data/SiggraphAsia2026VVC/selfcap/compression/test/0525_corgi.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/selfcap/compression/test/0811_yoga.tar | tar -xf - -C data/SiggraphAsia2026VVC/selfcap/compression/test/0811_yoga.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/selfcap/sparse/test/0512_bike.tar | tar -xf - -C data/SiggraphAsia2026VVC/selfcap/sparse/test/0512_bike.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/selfcap/sparse/test/0525_corgi.tar | tar -xf - -C data/SiggraphAsia2026VVC/selfcap/sparse/test/0525_corgi.tar
+curl -fL -H "Authorization: Bearer $(hf auth token)" https://huggingface.co/datasets/zju3dv/SiggraphAsia2026VVC/resolve/main/selfcap/sparse/test/0811_yoga.tar | tar -xf - -C data/SiggraphAsia2026VVC/selfcap/sparse/test/0811_yoga.tar
+```
+
+Each archive starts with the scene directory, for example `004_1_seq1/images/`.
+`checksums.sha256` lists the uploaded archives; those tar files are not kept.
+
+The held-out package is small and has camera YAML and no images.
+`heldout_test_cameras/checksums.sha256` checks that archive. Extract it so
+`test_intri.yml` and `test_extri.yml` land in each `.tar` scene directory.
+List the archive first. The transforms drop a leading `SiggraphAisa2026VVC/`
+prefix when it is present, and insert `.tar` before the scene directory, so
+`004_1_seq1/test_intri.yml` lands at `004_1_seq1.tar/004_1_seq1/test_intri.yml`.
+
+```bash
+tar -tf data/SiggraphAsia2026VVC/heldout_test_cameras/SiggraphAisa2026VVC_heldout_test_cameras.tar | head
+tar -xf data/SiggraphAsia2026VVC/heldout_test_cameras/SiggraphAisa2026VVC_heldout_test_cameras.tar \
+  -C data/SiggraphAsia2026VVC \
+  --transform 's,^SiggraphAisa2026VVC/,,' \
+  --transform 's,\([^/]*\)/\([^/]*\)$,\1.tar/\1/\2,'
+```
+
+Hardlink every extracted file into the scene directory passed to `--path`.
+Remove only those scene directories first. The `.tar` directories stay.
+
+```bash
+rm -rf \
+  data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/test/006_1_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/test/007_0_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/test/008_2_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/test/011_0_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/val/001_1_seq0 \
+  data/SiggraphAsia2026VVC/siga/compression/val/012_0_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/004_1_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/006_1_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/007_0_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/009_0_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/011_0_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/val/001_1_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/val/012_0_seq0 \
+  data/SiggraphAsia2026VVC/selfcap/compression/test/0512_bike \
+  data/SiggraphAsia2026VVC/selfcap/compression/test/0525_corgi \
+  data/SiggraphAsia2026VVC/selfcap/compression/test/0811_yoga \
+  data/SiggraphAsia2026VVC/selfcap/sparse/test/0512_bike \
+  data/SiggraphAsia2026VVC/selfcap/sparse/test/0525_corgi \
+  data/SiggraphAsia2026VVC/selfcap/sparse/test/0811_yoga
+mkdir -p \
+  data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/test/006_1_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/test/007_0_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/test/008_2_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/test/011_0_seq1 \
+  data/SiggraphAsia2026VVC/siga/compression/val/001_1_seq0 \
+  data/SiggraphAsia2026VVC/siga/compression/val/012_0_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/004_1_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/006_1_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/007_0_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/009_0_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/test/011_0_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/val/001_1_seq0 \
+  data/SiggraphAsia2026VVC/siga/sparse/val/012_0_seq0 \
+  data/SiggraphAsia2026VVC/selfcap/compression/test/0512_bike \
+  data/SiggraphAsia2026VVC/selfcap/compression/test/0525_corgi \
+  data/SiggraphAsia2026VVC/selfcap/compression/test/0811_yoga \
+  data/SiggraphAsia2026VVC/selfcap/sparse/test/0512_bike \
+  data/SiggraphAsia2026VVC/selfcap/sparse/test/0525_corgi \
+  data/SiggraphAsia2026VVC/selfcap/sparse/test/0811_yoga
+cp -al data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1.tar/004_1_seq1/. data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1/
+cp -al data/SiggraphAsia2026VVC/siga/compression/test/006_1_seq1.tar/006_1_seq1/. data/SiggraphAsia2026VVC/siga/compression/test/006_1_seq1/
+cp -al data/SiggraphAsia2026VVC/siga/compression/test/007_0_seq1.tar/007_0_seq1/. data/SiggraphAsia2026VVC/siga/compression/test/007_0_seq1/
+cp -al data/SiggraphAsia2026VVC/siga/compression/test/008_2_seq1.tar/008_2_seq1/. data/SiggraphAsia2026VVC/siga/compression/test/008_2_seq1/
+cp -al data/SiggraphAsia2026VVC/siga/compression/test/011_0_seq1.tar/011_0_seq1/. data/SiggraphAsia2026VVC/siga/compression/test/011_0_seq1/
+cp -al data/SiggraphAsia2026VVC/siga/compression/val/001_1_seq0.tar/001_1_seq0/. data/SiggraphAsia2026VVC/siga/compression/val/001_1_seq0/
+cp -al data/SiggraphAsia2026VVC/siga/compression/val/012_0_seq0.tar/012_0_seq0/. data/SiggraphAsia2026VVC/siga/compression/val/012_0_seq0/
+cp -al data/SiggraphAsia2026VVC/siga/sparse/test/004_1_seq0.tar/004_1_seq0/. data/SiggraphAsia2026VVC/siga/sparse/test/004_1_seq0/
+cp -al data/SiggraphAsia2026VVC/siga/sparse/test/006_1_seq0.tar/006_1_seq0/. data/SiggraphAsia2026VVC/siga/sparse/test/006_1_seq0/
+cp -al data/SiggraphAsia2026VVC/siga/sparse/test/007_0_seq0.tar/007_0_seq0/. data/SiggraphAsia2026VVC/siga/sparse/test/007_0_seq0/
+cp -al data/SiggraphAsia2026VVC/siga/sparse/test/009_0_seq0.tar/009_0_seq0/. data/SiggraphAsia2026VVC/siga/sparse/test/009_0_seq0/
+cp -al data/SiggraphAsia2026VVC/siga/sparse/test/011_0_seq0.tar/011_0_seq0/. data/SiggraphAsia2026VVC/siga/sparse/test/011_0_seq0/
+cp -al data/SiggraphAsia2026VVC/siga/sparse/val/001_1_seq0.tar/001_1_seq0/. data/SiggraphAsia2026VVC/siga/sparse/val/001_1_seq0/
+cp -al data/SiggraphAsia2026VVC/siga/sparse/val/012_0_seq0.tar/012_0_seq0/. data/SiggraphAsia2026VVC/siga/sparse/val/012_0_seq0/
+cp -al data/SiggraphAsia2026VVC/selfcap/compression/test/0512_bike.tar/0512_bike/. data/SiggraphAsia2026VVC/selfcap/compression/test/0512_bike/
+cp -al data/SiggraphAsia2026VVC/selfcap/compression/test/0525_corgi.tar/0525_corgi/. data/SiggraphAsia2026VVC/selfcap/compression/test/0525_corgi/
+cp -al data/SiggraphAsia2026VVC/selfcap/compression/test/0811_yoga.tar/0811_yoga/. data/SiggraphAsia2026VVC/selfcap/compression/test/0811_yoga/
+cp -al data/SiggraphAsia2026VVC/selfcap/sparse/test/0512_bike.tar/0512_bike/. data/SiggraphAsia2026VVC/selfcap/sparse/test/0512_bike/
+cp -al data/SiggraphAsia2026VVC/selfcap/sparse/test/0525_corgi.tar/0525_corgi/. data/SiggraphAsia2026VVC/selfcap/sparse/test/0525_corgi/
+cp -al data/SiggraphAsia2026VVC/selfcap/sparse/test/0811_yoga.tar/0811_yoga/. data/SiggraphAsia2026VVC/selfcap/sparse/test/0811_yoga/
+```
+
+Link images and masks, then write COLMAP text models. The default run starts
+at the first frame file and writes `frame1`. `siga` frames begin at
+`000000.jpg`. `selfcap` releases `000900.jpg` through `001799.jpg`, and
+`000900.jpg` is still `frame1`.
+
+```bash
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/compression/test/006_1_seq1
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/compression/test/007_0_seq1
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/compression/test/008_2_seq1
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/compression/test/011_0_seq1
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/compression/val/001_1_seq0
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/compression/val/012_0_seq0
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/sparse/test/004_1_seq0
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/sparse/test/006_1_seq0
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/sparse/test/007_0_seq0
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/sparse/test/009_0_seq0
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/sparse/test/011_0_seq0
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/sparse/val/001_1_seq0
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/siga/sparse/val/012_0_seq0
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/selfcap/compression/test/0512_bike
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/selfcap/compression/test/0525_corgi
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/selfcap/compression/test/0811_yoga
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/selfcap/sparse/test/0512_bike
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/selfcap/sparse/test/0525_corgi
+python -m nvs2colmap.SiggraphAsia2026VVC --path data/SiggraphAsia2026VVC/selfcap/sparse/test/0811_yoga
+```
+
+## Dataset Format
+
+A scene folder contains one directory of frames per view, EasyVolCap camera
+YAML, and optional masks and point clouds:
+
+```text
+004_1_seq1/
+  images/
+    00/000000.jpg
+    00/000001.jpg
+    01/000000.jpg
+    ...
+  masks/
+    00/000000.png
+    00/000001.png
+    ...
+  train_intri.yml
+  train_extri.yml
+  test_intri.yml
+  test_extri.yml
+  intri.yml
+  extri.yml
+  t_offsets.json
+  pcds/
+    000000.ply
+```
+
+Frames are `.jpg`. Output folders count those files from `frame1`, so the
+first file is `frame1` whether its name is `000000.jpg` or `000900.jpg`.
+
+`train_intri.yml` and `train_extri.yml` are the training views.
+`test_intri.yml` and `test_extri.yml` are the test views, and are read when
+those files are present. A view without `images/<view>/` is left out. A view
+listed in both YAML files keeps the training record. Views are sorted by name,
+so `images/00/000000.jpg` becomes `00.jpg`.
+
+Each YAML pair stores:
+
+- `names`: view ids, such as `00` and `01`.
+- `K_<view>`: 3 x 3 intrinsic matrix. `fx`, `fy`, `cx`, and `cy` are read from
+  it.
+- `H_<view>`, `W_<view>`: image height and width.
+- `R_<view>`: Rodrigues rotation of the world-to-camera pose.
+- `T_<view>`: world-to-camera translation. The pose is already in the COLMAP
+  camera convention.
+- `D_<view>`: OpenCV distortion coefficients `k1 k2 p1 p2 k3`. An all-zero
+  vector is left out. Any other value is written as COLMAP `OPENCV` using
+  `k1 k2 p1 p2`, and COLMAP undistorts `input/` into `images/`. The saved
+  camera is the resulting `PINHOLE` model.
+
+`t_offsets.json` stores a per-view time offset. The converter leaves it out of
+the COLMAP files. `pcds/000000.ply`, when present, is hardlinked to
+`sparse/0/points3D.ply` in every extracted frame.
+
+## Output Format
+
+In the default mode, each frame directory contains images, any masks, and a
+COLMAP text model. Zero-distortion images and masks are hardlinks. A scene with
+nonzero distortion also keeps the original files:
+
+```text
+0512_bike/
+  frame1/
+    input/
+      0002.jpg
+      ...
+    images/
+      0002.jpg
+      ...
+    input_mask/
+      0002.jpg.png
+      ...
+    image_masks/
+      0002.jpg.png
+      ...
+    sparse/0/
+      cameras.txt
+      images.txt
+      points3D.txt
+      points3D.ply
+```
+
+`input/` is a hardlink of the source image. `images/` is the file written by
+COLMAP's `image_undistorter`. A zero-distortion scene has no `input/` in the
+default mode; its hardlinks are only in `images/` and `image_masks/`.
+Mask filenames keep the image filename and add `.png`, so `images/0002.jpg`
+pairs with `image_masks/0002.jpg.png`, and `input/0002.jpg` pairs with
+`input_mask/0002.jpg.png`. A view with no `masks/<view>/` file gets an image
+and no mask. When `pcds/` is missing, `points3D.ply` is omitted.
+`--use-colmap` leaves the reconstructed points in `distorted/sparse/0` and
+does not link `pcds/`.
+
+With `--use-colmap`, images are linked into `frame*/input` and masks into
+`frame*/input_mask` (`input/0002.jpg` pairs with `input_mask/0002.jpg.png`).
+Each frame then runs feature extraction, matching, triangulation, mapping,
+`image_undistorter`, and `mask_undistorter`. Undistorted images are written to
+`images/`, and undistorted masks to `image_masks/`.
+
+All generated camera models use `PINHOLE`. Every camera gets its own COLMAP
+camera ID because these scenes are multi-view captures with known poses and
+per-view intrinsics. IDs start at 1 in view-name order.
+
+## Usage
+
+Link images and masks, then write COLMAP text models:
+
+```bash
+python -m nvs2colmap.SiggraphAsia2026VVC \
+  --path data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1
+```
+
+Start at source frame `000009.jpg` and keep output frame numbering aligned
+with it:
+
+```bash
+python -m nvs2colmap.SiggraphAsia2026VVC \
+  --path data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1 \
+  --start-number 10 \
+  --n-frames 20
+```
+
+`selfcap` sparse scenes such as `0512_bike` have no `masks/` and no
+`test_*.yml`. The released frames are `000900.jpg` through `001799.jpg`, and
+the default command writes them as `frame1` through `frame900`:
+
+```bash
+python -m nvs2colmap.SiggraphAsia2026VVC \
+  --path data/SiggraphAsia2026VVC/selfcap/sparse/test/0512_bike
+```
+
+Extract only the training cameras:
+
+```bash
+python -m nvs2colmap.SiggraphAsia2026VVC \
+  --path data/SiggraphAsia2026VVC/siga/compression/val/001_1_seq0 \
+  --no-test-camera
+```
+
+Run the full COLMAP pipeline for each frame:
+
+```bash
+python -m nvs2colmap.SiggraphAsia2026VVC \
+  --path data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1 \
+  --n-frames 1 \
+  --use-colmap \
+  --colmap-executable "${CONDA_EXE%/bin/conda}/bin/colmap" \
+  --colmap-use-gpu 1
+```
+
+Reuse existing linked frame images:
+
+```bash
+python -m nvs2colmap.SiggraphAsia2026VVC \
+  --path data/SiggraphAsia2026VVC/siga/compression/test/004_1_seq1 \
+  --skip-frame-linking
+```
+
+## Important Options
+
+- `--path`: scene directory containing `images/`, `train_intri.yml`, and
+  `train_extri.yml`.
+- `--n-frames`: number of frames to extract and write. If omitted, every frame
+  from `--start-number` through the last frame on disk is used. If omitted
+  with `--skip-frame-linking`, frame directories are counted from `frame%d`.
+- `--no-train-camera`: skip cameras listed in `train_intri.yml` and
+  `train_extri.yml`.
+- `--no-test-camera`: skip cameras listed in `test_intri.yml` and
+  `test_extri.yml`.
+- `--start-number`: 1-based position in the frames on disk, default `1`.
+  Output `frame%d` uses that same count. The first file is `frame1`.
+- `--skip-frame-linking`: keep existing frame images and masks, and skip
+  hardlinking and the distortion undistortion. A zero-distortion scene reuses
+  `frame*/images`, or `frame*/input` with `--use-colmap`. A scene with nonzero
+  distortion reuses `frame*/input` and the `images/` already written by COLMAP.
+- `--use-colmap`: run the full COLMAP pipeline instead of only writing
+  `sparse/0` text models.
+- `--colmap-executable`: path to the COLMAP executable. Nonzero distortion uses
+  it for `image_undistorter`, and `--use-colmap` uses it for the full pipeline.
+- `--colmap-use-gpu`: whether COLMAP SIFT extraction and matching should use
+  GPU, default `1`.

{nvs2colmap-0.7.0 → nvs2colmap-0.7.3}/nvs2colmap/SiggraphAsia2026VVC/__main__.py

@@ -9,7 +9,7 @@ from nvs2colmap.colmap import run_video_colmap
 from nvs2colmap.write_model import write_video_colmap_text_model
 
 from .cameras import read_camera_meta
-from .link_frames import count_frame_dirs, first_frame_number, link_frames, select_frame_indices
+from .link_frames import count_frame_dirs, link_frames, select_frame_indices
 from .points import link_point_clouds
 from .undistort import undistort_frames
 
@@ -38,10 +38,8 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "--start-number",
         type=int,
-        help=(
-            "1-based source frame to start from. Defaults to the first frame file. "
-            "Source file 000000.jpg is frame 1, and 000900.jpg is frame 901."
-        ),
+        default=1,
+        help="1-based position in the frames on disk. The first file is frame1.",
     )
     parser.add_argument(
         "--no-train-camera",
@@ -85,7 +83,7 @@ def parse_args() -> argparse.Namespace:
         parser.error("Cannot set both --no-train-camera and --no-test-camera.")
     if args.n_frames is not None and args.n_frames < 1:
         parser.error("--n-frames must be at least 1.")
-    if args.start_number is not None and args.start_number < 1:
+    if args.start_number < 1:
         parser.error("--start-number must be at least 1.")
     return args
 
@@ -99,39 +97,37 @@ def main() -> None:
         include_train=not args.no_train_camera,
         include_test=not args.no_test_camera,
     )
-    if args.skip_frame_linking:
-        start_number = args.start_number or first_frame_number(folder)
-        n_frames = args.n_frames
-        if n_frames is None:
-            n_frames = count_frame_dirs(folder / "frame%d", start_number=start_number)
-        frame_indices = list(range(start_number - 1, start_number - 1 + n_frames))
-    else:
-        frame_indices = select_frame_indices(available_frames, args.start_number, args.n_frames)
+    start_number = args.start_number
+    n_frames = args.n_frames
+    if args.skip_frame_linking and n_frames is None:
+        n_frames = count_frame_dirs(folder / "frame%d", start_number=start_number)
+    frame_indices = select_frame_indices(available_frames, start_number, n_frames)
+    output_numbers = list(range(start_number, start_number + len(frame_indices)))
 
     frame_output_pattern = folder / "frame%d"
     distorted = any(camera.dist is not None for camera in cameras)
-    image_dir_name = "input" if args.use_colmap else "images"
-    mask_dir_name = "input_mask" if args.use_colmap else "image_masks"
+    image_dir_name = "input" if args.use_colmap or distorted else "images"
+    mask_dir_name = "input_mask" if args.use_colmap or distorted else "image_masks"
     if not args.skip_frame_linking:
         link_frames(
             folder=folder,
             frame_pattern=frame_output_pattern,
             cameras=cameras,
             frame_indices=frame_indices,
+            output_numbers=output_numbers,
             image_dirname=image_dir_name,
             mask_dirname=mask_dir_name,
         )
 
     colmap_cameras = [camera.camera for camera in cameras]
     n_frames = len(frame_indices)
-    start_number = frame_indices[0] + 1
     if not args.use_colmap:
         if distorted:
             if not args.skip_frame_linking:
                 undistort_frames(
                     frame_pattern=frame_output_pattern,
                     cameras=cameras,
-                    frame_indices=frame_indices,
+                    output_numbers=output_numbers,
                     colmap_executable=args.colmap_executable,
                 )
         else:
@@ -145,7 +141,7 @@ def main() -> None:
         link_point_clouds(
             folder=folder,
             output_pattern=frame_output_pattern / "sparse" / "0" / "points3D.ply",
-            frame_indices=frame_indices,
+            output_numbers=output_numbers,
         )
     else:
         run_video_colmap(