SDGMNet: Statistic-Based Dynamic Gradient Modulation for Local Descriptor Learning

Authors: Yuxin Deng, Jiayi Ma

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our novel descriptor surpasses previous state-of-the-art methods in several tasks including patch verification, retrieval, pose estimation, and 3D reconstruction.
Researcher Affiliation Academia Yuxin Deng, Jiayi Ma Electronic Information School, Wuhan University, Wuhan 430072, China dyx_acuo@whu.edu.cn, jyma2010@gmail.com
Pseudocode Yes Algorithm 1: SDGMNet for local descriptor learning Input: m and α, initial β0, model, dataset, optimizer. t = 1; while training do Sample a data batch from datasets; Obtain Hard Net triplets in batch; Update corresponding statistics by Eq. (14); Compute weights by Eqs. (9), (10), and (11); if warming then Set w+ s (θ+), w s (θ ) and wc(θr) to 1 end if Compute powers P + and P by Eq. (12); Update the expectation of powers by Eq. (14); Construct pseudo loss by Eq. (15); Update the model with the optimizer; t = t + 1; end while Output: Well-trained model.
Open Source Code Yes Our codes are available at https://github.com/ACu Oo Oo O/SDGMNet.
Open Datasets Yes We train SGMNet on UBC Photo Tour dataset (Winder and Brown 2007)
Dataset Splits Yes UBC Photo Tour (Winder and Brown 2007) is the most widely used dataset for local descriptor learning. It consists of three subsets Liberty, Yosemite and Notredame. Deep descriptors are trained on one subset and tested on the other two. and The best checkpoint with the highest validation accuracy is chosen for the test.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory amounts) are mentioned for running experiments.
Software Dependencies No No specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions) are mentioned.
Experiment Setup Yes We train SGMNet on UBC Photo Tour dataset (Winder and Brown 2007) with Algorithm 1. where we set m = 0.6 and α = 0.9 for the best performance. The network is trained for 200 epochs (200K iterations) with batch size of 1024 and SGD optimizer. Moreover, the training is warmed up with w = 1 in the first 10% of iterations.