Uploading new results: neural-pbir

[sunset1995]

Thanks for publishing the awesome datasets! I want to upload the results by our ICCV 2023 paper --- Neural-PBIR Reconstruction of Shape, Material, and Illumination. Below are the detailed of our results:

• model name: neural-pbir
• paper link: https://arxiv.org/abs/2304.13445
• website link: https://neural-pbir.github.io/
• benchmarks: full
• results: https://drive.google.com/drive/folders/1xiN_tSZgfDLM-odIGbeIavwQQXKY0VNY?usp=sharing
• eval input json: neural-pbir-eval-inputs.json
• eval output json: neural-pbir-eval-outputs.json
  • I run the following script to get the evaluation results: python -m scripts.test --input-path "$folder"/neural-pbir-eval-inputs.json --output-path "$folder"/neural-pbir-eval-outputs.json --scenes full

[zfkuang]

Hi Cheng,

Thanks for the update and nice work! I've verified the scores and uploaded them to the website now.

Best, Zhengfei

[sunset1995]

Hi Zhengfei,

Thanks for organizing the benchmark!

One question about the shape results. The chamfer distance from my evaluation result is 0.214 * 1e-3 but it appears to be 0.43 * 1e-3 on the website table. Do I misinterpret the result? If it is, I have to fix it in our arxiv version as well.

        "shape": {
            "bidir_chamfer": [
                0.0002145684793258237,
                0.0002457779565475051
            ],
            "scene_count": 42
        }

Best, Cheng

[zfkuang]

Hi Cheng,

In fact, we multiplied the chamfer distances with 2x10^3 to get the table numbers, which is a bit different from what we stated in the paper. Sorry for the confusion. We'll update this repo's document very soon.

Best, Zhengfei

[sunset1995]

Got it. I will correct the number in our arxiv version.