VA-DepthNet: A Variational Approach to Single Image Depth Prediction

Authors: Ce Liu, Suryansh Kumar, Shuhang Gu, Radu Timofte, Luc Van Gool

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

Reproducibility Variable Result LLM Response
Research Type Experimental The paper demonstrates the usefulness of the proposed approach via extensive evaluation and ablation analysis over several benchmark datasets, such as KITTI, NYU Depth V2, and SUN RGB-D.
Researcher Affiliation Academia 1CVL ETH Zürich 2UESTC China 3University of Würzburg 4KU Leuven
Pseudocode No No pseudocode or clearly labeled algorithm blocks were found in the paper.
Open Source Code Yes For official code refer here
Open Datasets Yes Datasets. We performed experiments on three benchmark datasets namely NYU Depth V2 (Silberman et al., 2012), KITTI (Geiger et al., 2012), and SUN RGB-D (Song et al., 2015).
Dataset Splits Yes (b) KITTI contains images with 352 1216 resolution where depth values range from 0 to 80 meters. The official split provides 42,949 train, 1,000 validation, and 500 test images. Eigen et al. (2014) provides another train and test set split for this dataset which has 23,488 train and 697 test images.
Hardware Specification Yes We implemented our method in Py Torch 1.7.1 (Python 3.8) with CUDA 11.0. The software is evaluated on a computing machine with Quadro-RTX-6000 GPU.
Software Dependencies Yes We implemented our method in Py Torch 1.7.1 (Python 3.8) with CUDA 11.0.
Experiment Setup Yes We use the Adam optimizer (Kingma & Ba, 2014) without weight decay. We decrease the learning rate from 3e 5 to 1e 5 by the cosine annealing scheduler. To avoid over-fitting, we augment the images by horizontal flipping. For KITTI (Geiger et al., 2012), the model is trained for 10 epochs for the official split and 20 epochs for the Eigen split (Eigen et al., 2014). For NYU Depth V2 (Silberman et al., 2012), the model is trained for 20 epochs.