DCDepth: Progressive Monocular Depth Estimation in Discrete Cosine Domain

Authors: Kun Wang, Zhiqiang Yan, Junkai Fan, Wanlu Zhu, Xiang Li, Jun Li, Jian Yang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct comprehensive experiments on NYU-Depth V2, TOFDC, and KITTI datasets, and demonstrate the state-of-the-art performance of DCDepth.
Researcher Affiliation Academia 1PCA Lab, Nanjing University of Science and Technology, China 2Nankai University, China
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/w2kun/DCDepth.
Open Datasets Yes We evaluate our method on three datasets that covers a diverse array of indoor and outdoor scenes. (1) NYU-Depth-V2 [36]... (2) TOFDC [52]... (3) KITTI [13]...
Dataset Splits No We follow the data split as outlined in BTS [17], featuring 24231 training images and 654 test images. The dataset is divided into 10,000 training samples and 560 testing samples. The Eigen split comprises 23158 training images and 697 test images, while the official split includes 42949 training images and 500 test images. The paper provides specific train and test splits but does not explicitly mention validation splits for all datasets.
Hardware Specification Yes is trained with a batch size of 8 on four NVIDIA RTX-4090 GPUs with data-distributed parallel computing.
Software Dependencies No The DCDepth is implemented using Pytorch library [25] (No specific version number for PyTorch or other software dependencies is provided).
Experiment Setup Yes The DCDepth is implemented using Pytorch library [25], and is trained with a batch size of 8 on four NVIDIA RTX-4090 GPUs... Our method is trained on NYU-Depth-V2 dataset for 20 epochs... The optimization objective of our method is a combination of the scale-invariant log loss Ld, the frequency regularization Lf and the smoothness regularization Ls, weighted by two scalar weights α and β... We opt for the Adam optimizer [16] and leverage the One Cycle learning rate scheduler [37].