Multi-task Learning with 3D-Aware Regularization

Authors: Wei-Hong Li, Steven McDonagh, Ales Leonardis, Hakan Bilen

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

Reproducibility Variable Result LLM Response
Research Type Experimental We show that the proposed method is architecture agnostic and can be plugged into various prior multi-task backbones to improve their performance; as we evidence using standard benchmarks NYUv2 and PASCAL-Context.
Researcher Affiliation Academia 1University of Edinburgh, 2University of Birmingham
Pseudocode No The paper does not contain any explicit pseudocode or algorithm blocks.
Open Source Code Yes github.com/VICO-Uo E/3DAware MTL
Open Datasets Yes NYUv2 (Silberman et al., 2012): It contains 1449 RGB-D images...PASCAL-Context (Chen et al., 2014): PASCAL (Everingham et al., 2010) is a commonly used image benchmark for dense prediction tasks.
Dataset Splits No The paper mentions following 'identical training, evaluation protocols' from other works (Ye & Xu, 2022) but does not explicitly provide the specific percentages or counts for train/validation/test splits within the paper itself.
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU or CPU models. It only refers to 'standard GPUs' generally when discussing memory limitations.
Software Dependencies No The paper mentions using frameworks and models like MTI-Net, Inv PT, HRNet-48, and ViT-L, but it does not specify the version numbers of any software dependencies (e.g., PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes We append our 3D-aware regularizer to MTI-Net and Inv PT using two convolutional layers, followed by Batch Norm, ReLU, and dropout layer with a dropout rate of 0.15...We train all models for 40K iterations with a batch size of 6 for experiments of using Inv PT...and a batch size of 8 for experiments of using MTI-Net...We ramp up the αt from 0 to 4 linearly in 20K iterations and keep αt = 4 for the rest 20K iterations.