DenserNet: Weakly Supervised Visual Localization Using Multi-Scale Feature Aggregation

Authors: Dongfang Liu, Yiming Cui, Liqi Yan, Christos Mousas, Baijian Yang, Yingjie Chen6101-6109

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiment results indicate that our method sets a new state-of-the-art on four challenging large-scale localization benchmarks and three image retrieval benchmarks with the same level of supervision.
Researcher Affiliation Academia 1Purdue University 2University of Florida 3Fudan University
Pseudocode No The architecture of Denser Net is shown in Figure 2. Denser Net includes a CNN backbone, three feature extraction branches, and a one-stage feature decoder.
Open Source Code Yes The code is available at https://github.com/goodproj13/ Denser Net.
Open Datasets Yes We train the proposed method by using Pitts30k-training dataset (Arandjelovic et al. 2016). Following (Arandjelovic et al. 2016; Jin Kim, Dunn, and Frahm 2017), we group the positive {I+ r } and negative {I r } images for each training query image It.
Dataset Splits Yes The 30K training query images generate four image triplets. Thus, we obtain a total of 120K image triplets with 112K for training and 8K for validation.
Hardware Specification Yes All experiments are performed on a workstation with an Intel Core i7-7820X CPU and four NVIDIA Ge Force GTX 3080Ti GPU.
Software Dependencies No Both VGG16 and Mobile Net V2 based methods are pretrained on Image Net (Deng et al. 2009). In training, we exploit standard data augmentation in training, such as motion blur, random Gaussian noise, brightness changes to improve the robustness of our methods to illumination variations and viewpoint changes. Specifically, the margin M is set at 0.1, 30 epochs are performed using batch size of 4 triplets, Adam (Kingma and Ba 2014) with the learning rates of 10 3 which is halved every 6 epochs, momentum of 0.9, and weight decay of 10 3.
Experiment Setup Yes Specifically, the margin M is set at 0.1, 30 epochs are performed using batch size of 4 triplets, Adam (Kingma and Ba 2014) with the learning rates of 10 3 which is halved every 6 epochs, momentum of 0.9, and weight decay of 10 3.