Inverse Rendering using Multi-Bounce Path Tracing and Reservoir Sampling
Authors: Yuxin Dai, Qi Wang, Jingsen Zhu, xi db, Yuchi Huo, Chen Qian, Ying He
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through both qualitative and quantitative assessments across various scenarios, especially those with complex shadows, we demonstrate that our method achieves state-of-the-art decomposition performance. Furthermore, our optimized explicit geometry seamlessly integrates with modern graphics engines supporting downstream applications such as scene editing, relighting, and material editing. |
| Researcher Affiliation | Collaboration | Yuxin Dai1 , Qi Wang2 , Jingsen Zhu3 , Dianbing Xi2, Yuchi Huo2, Chen Qian4, Ying He1 1S-Lab, Nanyang Technological University 2State Key Lab of CAD&CG, Zhejiang University 3Cornell University 4Sense Time Research and Tetras.AI |
| Pseudocode | No | The paper describes the methodology and pipeline in detail across various sections (e.g., Section 3, Section 4) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | Project page: https://brabbitdousha.github.io/MIRRe S/. |
| Open Datasets | Yes | We evaluate our method on two benchmark datasets for inverse rendering: (1) Tenso IR synthetic dataset Jin et al. (2023), which provides ground-truth geometry, material parameters, and relighted images, and (2) Objects-with-Lighting (OWL) real dataset Ummenhofer et al. (2024), from which we select four scenes (Antman, Tpiece, Gamepad, and Porcelain Mug) containing ground-truth relighted images and environment maps. |
| Dataset Splits | No | The paper mentions evaluating on 'testing images' from benchmark datasets but does not explicitly provide specific percentages, counts, or detailed methodologies for how these splits were performed or sourced for the experiments. |
| Hardware Specification | Yes | We run our training and inference on a single NVIDIA RTX 4090 GPU, with the entire two-stage training process taking approximately 4.5 hours. |
| Software Dependencies | No | We implement MIRRe S using Pytorch framework Paszke et al. (2019) with CUDA extensions in SLANG.D Bangaru et al. (2023). We customize CUDA kernels in our rendering layer to perform efficient reservoir sampling and multi-bounce path tracing. We also utilize nvdiffrast Laine et al. (2020) for differentiable ray-mesh intersection. |
| Experiment Setup | Yes | Training follows the structure described in Section 3. The first stage is identical to the training of Instant NGP M uller et al. (2022) and Neu S2 Wang et al. (2023a), for which we refer to their papers for details such as training losses. Our model is trained by rendering loss (Eqs. (6) and (7)) along with several regularization terms, which are fully specified in the appendix. Network details. In stage 1, the density and appearance fields Fσ, Fc follow the standard Instant-NGP configuration M uller et al. (2022), using a hash grid with 16 levels, 2 feature dimensions per entry, a coarsest resolution of Nmin = 16, and a finest resolution of Nmax = 2048, followed by a 4-layer MLP with 64 hidden channels. In stage 2, the material network Fm uses the same hash grid configuration, followed by a 2-layer MLP with 32 hidden channels. Training details. Apart from the rendering losses (Eqs. (6) and (7)) mentioned in the main text, we also add several additional regularizations to stabilize the training, described as follows: To prevent drastic changes in vertex offset v during optimization, we apply the Laplacian smooth loss and vertices offset regularization loss from Ne RF2Mesh Tang et al. (2023b): ... We also apply the smoothness regularizers for albedo kd, roughness ρ, and normal n proposed by NVdiffrec-MC Hasselgren et al. (2022) for better intrinsic decomposition: ... Additionally, for better disentangling material parameters and light, we adopt the same monochrome regularization term of NVdiffrec-MC Hasselgren et al. (2022): ... Thus, we use 32 SPP as the default configuration for all our experiments in this section. |