Bugjudger/HDR_Pretrain

HDR-Pretrain: A Pretraining Dataset for 3D HDR Reconstruction

Paper | Code | Project page

HDR-Pretrain is a synthetic multi-view dataset for pretraining 3D high dynamic range (HDR) reconstruction models. It provides multi-exposure low dynamic range (LDR) images and corresponding HDR references to support learning 3D HDR scene representations from multi-view observations and rendering HDR images from novel viewpoints.

The dataset contains 168 Blender-rendered indoor scenes. Each scene contains 18 training views and 17 test views, with five LDR exposure variants per view, HDR references, camera metadata, depth, and normals. The source Blender scenes are also included for re-rendering and data generation.

Directory layout

HDR-Pretrain/
├── train_index.json
├── test_index.json
├── scenes/
│   └── 102343992/
│       ├── exposure.json
│       ├── images/
│       ├── images_hdr/
│       ├── sparse/0/
│       │   ├── cameras.bin
│       │   └── images.bin
│       ├── depth/
│       ├── normal/
│       ├── transforms_train.json
│       └── transforms_test.json
├── blender_scenes/
│   ├── 102343992.blend
│   └── ...
├── instanthdr_render.py
└── instanthdr.sh

Training and validation indexes

The indexes are JSON arrays of scene paths relative to this directory, for example:

["scenes/102343992", "scenes/102344022"]

train_index.json includes all 168 scenes. test_index.json contains five scenes selected with Python's random.Random(42).sample from the sorted scene list. These scenes also participate in training: the loader uses train views during training and test views during validation. This split does not measure generalization to unseen scenes.

The five validation scenes are:

• 102344250
• 102816786
• 104348328_171513363
• 104862501_172226556
• 108736800_177263517

Set dataset.hdr.roots to the directory containing this README, rather than its scenes/ subdirectory. Exposure JSON entries are positive linear exposure times. The sparse/0/ files contain Blender-exported camera metadata; no COLMAP reconstruction or sparse point cloud is required. The current training loader uses these files instead of transforms_*.json.

Generating scenes

Install Blender 4.5.6 and run the batch launcher, setting the executable path and GPU as needed:

BLENDER=/path/to/blender CUDA_VISIBLE_DEVICES=0 bash instanthdr.sh

The launcher reads .blend files from blender_scenes/ and writes rendered data into scenes/<scene_id>/. Rendering settings are in the CONFIG dictionary in instanthdr_render.py. The supplied indexes cover the existing 168 scenes; update the indexes if adding new scene IDs.