Discrete Cycle-Consistency Based Unsupervised Deep Graph Matching

Authors: Siddharth Tourani, Muhammad Haris Khan, Carsten Rother, Bogdan Savchynskyy

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental evaluation suggests that our technique sets a new state-of-the-art for unsupervised deep graph matching.
Researcher Affiliation Academia Siddharth Tourani1,2, Muhammad Haris Khan2, Carsten Rother1, Bogdan Savchynskyy1 1Computer Vision and Learning Lab, IWR, Heidelberg University, 2MBZUAI
Pseudocode Yes Algorithm 1: Unsupervised training algorithm
Open Source Code Yes Our code is available at: https://github.com/skt9/ clum-semantic-correspondence.
Open Datasets Yes Datasets We evaluate our proposed method on the task of keypoint matching on the following datasets: Willow Object Class (Cho, Alahari, and Ponce 2013), SPair-71K (Min et al. 2019) and Pascal VOC with Berkeley annotations (Everingham et al. 2010; Bourdev and Malik 2009).
Dataset Splits No The paper mentions evaluating on datasets like Pascal VOC, Willow, and SPair-71K but does not explicitly provide the specific percentages or sample counts for training, validation, and test splits within the paper. It mentions processing 'batches of 12 image triplets' for training but not the overall dataset splits.
Hardware Specification Yes All experiments were run on an Nvidia-A100 GPU and a 32 core CPU.
Software Dependencies No The paper mentions using specific network architectures like VGG16 and Spline CNN layers, and combinatorial solvers like LPMP and fusion moves. However, it does not provide specific version numbers for any software dependencies (e.g., Python, PyTorch, CUDA, or the solvers themselves).
Experiment Setup Yes The hyper-parameters are the same in all experiments. We used Adam (Kingma and Ba 2015) with an initial learning rate of 2 10 3 which is halved at regular intervals. The VGG16 backbone learning rate is multiplied by 0.01. We process batches of 12 image triplets. The hyper-parameter λ from (3) is set to 80. Hyper-parameter ˆc from (10) for Pascal VOC (unfiltered) is set to 0.21 for SCGM w/BBGM, 0.257 for both CLUM and CLUM-L, 0.329 for both CL-BBGM and CL-BBGM (SCGM), respectively. We use image flips and rotations as augmentations.