CoolFace
Datasetpublic

nnuochen/RGBT-Scenes-raw2temp-inferno-p0-98-filtered

RGBT-Scenes Raw2Temp Inferno Filtered This dataset is a full-copy variant of the local RGBT-Scenes extraction used by MM3DGS experiments. What Changed thermal/ is replaced with images rendered from FLIR raw_images/*.jpg using: RawThermalImage extraction radiometric raw2temp conversion percentile stretch p0-p98 inferno colormap resize to 640x480 The following three bad views are removed from the dataset entirely: Building/img_001.jpg Parterre/FLIR5559.jpg… See the full description on the dataset page: https://huggingface.co/datasets/nnuochen/RGBT-Scenes-raw2temp-inferno-p0-98-filtered.

sourceHugging Faceunknownupdated 6mo agoView on Hugging Face
0likes232downloads
Dataset Card

RGBT-Scenes Raw2Temp Inferno Filtered

This dataset is a full-copy variant of the local RGBT-Scenes extraction used by MM3DGS experiments.

What Changed

  • —thermal/ is replaced with images rendered from FLIR raw_images/*.jpg using:
  • —RawThermalImage extraction
  • —radiometric raw2temp conversion
  • —percentile stretch p0-p98
  • —inferno colormap
  • —resize to 640x480
  • —The following three bad views are removed from the dataset entirely:
  • —Building/img_001.jpg
  • —Parterre/FLIR5559.jpg
  • —Truck/FLIR3077.jpg
  • —This is a real copy of the dataset tree. No symbolic links are used.

Scene Layout

Each scene keeps the same top-level folders as the local extracted dataset:

  • —images/
  • —thermal/
  • —msx/
  • —raw_images/
  • —depths_da3_mv/
  • —depths_da3_mv_raw/
  • —depths_da3_sv/
  • —depths_da3_sv_raw/
  • —depths_official_vitl/
  • —sparse/0/
  • —colmap/sparse/0/

The filtered sparse/0 and colmap/sparse/0 models were rewritten so the removed views are no longer referenced by COLMAP images.bin.

Counts

  • —scenes: 10
  • —total remaining views: 1030
  • —removed views: 3

Per-scene counts and filtered-model statistics are recorded in filtered_dataset_manifest.json.

Notes

  • —colmap/database.db is copied from the source extraction and is not rewritten.
  • —MM3DGS training uses the filtered sparse/0 model and the scene image folders, so the dataset is directly usable for training/evaluation.