Learning to Annotate Part Segmentation with Gradient Matching

Authors: Yu Yang, Xiaotian Cheng, Hakan Bilen, Xiangyang Ji

ICLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our method is evaluated with semi-supervised part segmentation tasks and significantly outperforms other semi-supervised competitors when the amount of labelled examples is extremely limited.
Researcher Affiliation Academia Yu Yang Department of Automation Tsinghua University, BNRist EMAIL Xiaotian Cheng Department of Automation Tsinghua University, BNRist EMAIL Hakan Bilen School of Informatics University of Edinburgh EMAIL Xiangyang Ji Department of Automation Tsinghua University, BNRist EMAIL
Pseudocode Yes Algorithm 1: Learning to annotate with gradient matching. Algorithm 2: Learning to annotate with K-step MAML.
Open Source Code Yes Code is available at https://github.com/yangyu12/lagm.
Open Datasets Yes We evaluate our method on six part segmentation datasets: Celeb A, Pascal-Horse, Pascal Aeroplane, Car-20, Cat-16, and Face-34. Celeb A (Liu et al., 2015)... Pascal Part (Chen et al., 2014)... LSUN (Yu et al., 2015)... Car-20, Cat-16, and Face-34, released by Zhang et al. (2021)...
Dataset Splits Yes We finally obtain 180 training images, 34 validation images, and 225 test images to constitute Pascal-Horse, 180 training images, 78 validation images, and 269 test images to constitute Pascal-Aeroplane.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) are mentioned in the paper, only general terms like 'GPU memory'.
Software Dependencies No The paper mentions software like PyTorch, Deep Labv3, and U-Net, and links to a GitHub repository for Style GAN models, but does not provide specific version numbers for these software dependencies (e.g., 'PyTorch library' without a version).
Experiment Setup Yes The annotator and the segmentation network are optimized with an SGD optimizer with learning rate 0.001 and momentum 0.9. By default, we jointly train an annotator and a segmentation network with K = 1 and batch size 2 for 150,000 steps.