Differentiable Blocks World: Qualitative 3D Decomposition by Rendering Primitives

Authors: Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei Efros, Mathieu Aubry

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments
Researcher Affiliation Academia 1LIGM, Ecole des Ponts, Univ Gustave Eiffel 2UC Berkeley
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code and video results are available at www.tmonnier.com/DBW.
Open Datasets Yes DTU [25] is an MVS dataset containing 80 forward-facing scenes... We evaluate on 10 scenes (S24, S31, S40, S45, S55, S59, S63, S75, S83, S105)... The first row corresponds to the Campanile scene from Nerfstudio repository [66] and the last four rows correspond to Blended MVS scenes [75].
Dataset Splits No The paper describes using all available views for each scene (49 or 64 depending on the scenes) from the DTU dataset, but does not specify explicit training/validation/test dataset splits with percentages or sample counts.
Hardware Specification Yes Optimizing our model on a scene roughly takes 4 hours on a single NVIDIA RTX 2080 Ti GPU.
Software Dependencies No The paper mentions using PyTorch3D framework but does not provide specific version numbers for it or other key software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes Our model is optimized using Adam [31] with a batch size of 4 for roughly a total of 25k iterations. We use learning rates of 0.05 for the texture images and 0.005 for all other parameters, and divide them by 10 for the last 2k iterations.