Stanford-ORB: A Real-World 3D Object Inverse Rendering Benchmark

This is the official repository of the Stanford-ORB dataset to test 3D object-centric inverse rendering models.

The dataset consists of:

• 2,795 HDR images of 14 objects captured in 7 in-the-wild scenes;
• 418 HDR ground truth environment maps aligned with image captures;
• Textured scanned mesh of 14 objects captured from studio;
• Comprehensive benchmarks for evaluating the inverse rendering methods;
• Reported results of concurrent state-of-the-art models;
• A full set of scripts and guidelines to run the benchmarks.

This repository contains instructions for dataset downloads and evaluation tools.

Stanford-ORB: A Real-World 3D Object Inverse Rendering Benchmark
Zhengfei Kuang*, Yunzhi Zhang*, Hong-Xing Yu, Samir Agarwala, Shangzhe Wu, Jiajun Wu NeurIPS 2023 Datasets and Benchmarks Track, December 2023
Project Page / Paper

Dataset Structure

The dataset is provided under three commonly used representations with LDR and HDR versions:

1. The blender representation from the NeRF blender dataset (blender_HDR/LDR);
2. The LLFF representation (llff_colmap_HDR/LDR);
3. The COLMAP representation(llff_colmap_HDR/LDR).

Download links are provided on the project page.

data	
----- blender_HDR/LDR		
    ----- <obj_name>_scene00x               (name of the object and scene)
        ----- test/%04d.{exr/png}           (Test images in HDR/LDR)	
        ----- test_mask/%04d.png            (Test mask images in LDR)	
        ----- train/%04d.{exr/png}          (Train images in HDR/LDR)	
        ----- train_mask/%04d.png           (Train mask images in LDR)	
        ----- transforms_test.json          (Test metadata)	
        ----- transforms_train.json         (Train metadata)	
        ----- transforms_novel.json         (metadata of novel scene test images. Note that each frame has a unique camera angle)				
----- llff_colmap_HDR/LDR
    ----- <obj_name>_scene00x               (name of the object and scene)
        ----- images/%04d.{exr/png}         (All images in HDR/LDR)
        ----- masks/%04d.png                (All image masks in LDR)
        ----- sparse/0/*.bin                (Colmap Files)	
        ----- sparse/<nv_sn_name>/*.bin     (Colmap Files of novel scenes)	
        ----- poses_bounds.npy              (LLFF's camera files)	
        ----- train_id.txt                  (name of training images)	
        ----- test_id.txt                   (name of test images)	
        ----- novel_id.txt                  (name of test images from novel scenes)	
        ----- poses_bounds_novel.npy        (camera poses of novel scene test images indexed by order of names in novel_id.txt)					
----- ground_truth
    ----- <obj_name>_scene00x               (name of the object and scene)	
        ----- env_map/%04d.exr              (Ground truth HDRI environment maps (share the same name with the corresponding test image))
        ----- surface_normal/%04d.npy       (Ground truth surface normal maps)
        ----- z_depth/%04d.npy              (Ground truth Z-depth maps)
        ----- pseudo_gt_albedo/%04d.npy     (Pseudo GT albedo maps)
        ----- mesh_blender/mesh.obj         (Scanned Mesh aligned with the blender format data)

While most part of our data are identical to the original representations, some nuances still exist.

First, we provide object masks for all images, stored in the folder of *_mask for blender representation and masks for LLFF/Colmap representation. Feel free to use/ignore them during training.

Second, to support the task of relighting, in each data folder (e.g. baking_scene001), we also provide the camera poses of test images from other scenes (e.g. baking_scene002 and baking_scene003). In the blender representation, the data is stored in transforms_novel.json; In the LLFF representation it's stored in novel_id.txt and poses_bounds_novel.npy; In the COLMAP representation it's stored in sparse/<novel_scene_name>/*.bin. Note that all the novel camera poses are transformed to align with the original poses, so no further adaptation is required for the users. In other words, you can directly use them as additional test poses as shown in the example dataloader here.

Quick Start

With the dataset downloaded, you can train/test your model within only a few steps:

1. Training

We provide the example dataloaders for all structures here. Select one of the data dataloaders that best fits your method, and integrate it into your code.

Note: For accurate evaluation, your model must be trained by each capture separately (42 times in total).

2.Inference

Once your model is trained with one capture (e.g. baking_scene001), evaluation is done with the test data from the same capture (denoted as test dataset below) and the data from other captures of the same object, such as baking_scene002 and baking_scene003 (denoted as novel datasets below).

Here we explain the input and output of our benchmark tests:

• Novel View Synthesis
  • Input 1: Poses from the test dataset;
  • Output 1: Rendered LDR/HDR images of the test views. (in .jpg/.png/.exr format)
• Relighting:
  • Input 1: Poses from novel datasets;
  • Input 2: Ground truth environment maps of the test views (located in ground_truth/<capture_name>/env_map);
  • Output 1: Rendered LDR/HDR images of the test views. (in .jpg/.png/.exr format)
• Geometry Estimation:
  • Input 1: Poses from the test dataset;
  • Output 1: Rendered Z-depth maps of the test views. (in .npy format)
  • Output 2: Rendered camera-space normal maps of the test views. (in .npy format)
  • Output 3: Reconstructed 3D mesh. (in .obj format)

Run your model trained with each capture to get the required outputs, and then pack up the paths to your predictions and the corresponding ground truths in a json file. To do so, you may adapt this example script and run

python scripts/test_cache.py --method mymethod --output_path <path_to_test_results_json_file> 

A json file with the same structure as this example will be generated .

3. Evaluation

Evaluation requires the following conda environment:

# Tested on python 3.9 or above
conda create -n orb python=3.9
conda activate orb
# Install PyEXR
pip install git+https://github.com/jamesbowman/openexrpython.git (dependency) # or try `conda install -c conda-forge openexr-python`
pip install git+https://github.com/tvogels/pyexr.git
# Install other packages
pip install tqdm PyEXR trimesh lpips kornia
pip install scipy opencv-python matplotlib imageio[ffmpeg]

Simply run this one-line command:

python scripts/test.py --input-path <path_to_test_results_json_file> --output-path <path_to_test_scores_json_file> --scenes full

Ta-da! All evaluation results (summed-up scores and per-capture scores) will be written to the output json file.

Important Note: In the paper we stated that the chamfer distances are multiplied by $10^3$ in the table. However, the table numbers are actually $2 \times 10^3$ times the numbers in the json file. Sorry for confusion and please make sure to use the right scale.

4. Result Visualization

To draw visualization figures as in our paper, first, install the following required packages:

pip install numpy matplotlib glob2 tqdm seaborn==0.12.2 pandas==1.4.4 scipy==1.9.1 

Then, move the output json file to visulize/methods/<your_method_name>.json. Add your model's name in visulize/visualize.ipynb and run the script to draw the figure.

Feel free to play with the script for a better look.

5. How to upload your scores/results

If you'd like to upload your results and scores to our website, please open an issue in this repo with the title "Uploading new results: <your model name>", and provide the following informations:

• Name of your model, the paper/website link
• The selection of the benchmarks that you'd like to test;
• All testing results from your model (including 3D mesh, relit/novel view images (LDR and HDR) and depth/normal maps. Albedo maps are not required);
• A json file similar to this example, containing the paths to the result files;
• [Optional] The evaluation scores you get.
• [Optional] If your model runs with fewer training images than usual, you can also report the number of views in your submission. We will check that information in the paper and add it to the leaderboard as well.

After receiving your request, we will verify the scores shortly and update your scores to our website.

Citation

@misc{kuang2023stanfordorb,
      title={Stanford-ORB: A Real-World 3D Object Inverse Rendering Benchmark}, 
      author={Zhengfei Kuang and Yunzhi Zhang and Hong-Xing Yu and Samir Agarwala and Shangzhe Wu and Jiajun Wu},
      year={2023},
      eprint={2310.16044},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}