To Choose or to Fuse? Scale Selection for Crowd Counting

Authors: Qingyu Song, Changan Wang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Jian Wu, Jiayi Ma2576-2583

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

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the superiority of our method, we conduct extensive experiments on four challenging datasets, including Shanghai Tech dataset (Zhang et al. 2016), UCF CC 50 dataset (Idrees et al. 2013), UCF-QNRF dataset (Idrees et al. 2018) and World Expo 10 dataset (Zhang et al. 2015).
Researcher Affiliation Collaboration 1Tencent Youtu Lab, Shanghai, China 2College of Computer Science & Technology, Zhejiang University, Hangzhou, China 3Electronic Information School, Wuhan University, Wuhan, China
Pseudocode No The paper describes algorithms and processes (e.g., "(a) Searching Algorithm"), but it does not present any formal pseudocode blocks or figures explicitly labeled as "Pseudocode" or "Algorithm".
Open Source Code Yes The code will be available at: https://github.com/Tencent Youtu Research/Crowd Counting SASNet.
Open Datasets Yes We conduct extensive experiments on four challenging datasets, including Shanghai Tech dataset (Zhang et al. 2016), UCF CC 50 dataset (Idrees et al. 2013), UCF-QNRF dataset (Idrees et al. 2018) and World Expo 10 dataset (Zhang et al. 2015).
Dataset Splits Yes We follow previous work (Idrees et al. 2013) to conduct a five-fold cross validation.
Hardware Specification No The paper describes the model architecture and training process but does not specify any hardware details such as GPU models, CPU types, or memory used for experiments.
Software Dependencies No The codes are implemented using the open-source C3-framework (Gao et al. 2019). We optimize the model using Adam algorithm (Kingma and Ba 2014).
Experiment Setup Yes During training, we randomly select eight images for each epoch and crop four image patches with a fixed-size of 128 128 from each image. Thus the effective batch size is 32. We also use random horizontal flipping with a probability of 0.5 as data augmentation. The image patch size k in the confidence branch is set as 32. The weight term γ is set as 1 and the λ is set to 10. We optimize the model using Adam algorithm (Kingma and Ba 2014) with a fixed learning rate of 1e-5.