UCLID-Net: Single View Reconstruction in Object Space

Authors: Benoit Guillard, Edoardo Remelli, Pascal Fua

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate both on Shape Net synthetic images, which are often used for benchmarking purposes, and on real-world images that our approach outperforms state-of-the-art ones.
Researcher Affiliation Academia Benoit Guillard Edoardo Remelli CVLab EPFL, Switzerland {firstname.lastname}@epfl.ch
Pseudocode No The paper describes the architecture and processes through textual descriptions and diagrams but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available here: https://github.com/cvlab-epfl/UCLID-Net.
Open Datasets Yes Shape Net Core [2] features 38000 shapes belonging 13 object categories... PIX3D [19] is a collection of pairs of real images of furniture with ground truth 3D models and pose annotations.
Dataset Splits Yes We use the same testing and training splits but re-generated the depth maps because the provided ones are clipped along the z-axis. We therefore use it for validation only, on approximately 2.5k images of chairs.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU/CPU models, memory).
Software Dependencies No The paper mentions 'implemented in Pytorch' but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes The network is implemented in Pytorch, and trained for 150 epochs using the Adam optimizer, with initial learning rate 10 3, decreased to 10 4 after 100 epochs.