Localization in the Crowd with Topological Constraints
Authors: Shahira Abousamra, Minh Hoai, Dimitris Samaras, Chao Chen872-881
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On multiple public benchmarks, our method outperforms previous localization methods. Additionally, we demonstrate the potential of our method in improving the performance in the crowd counting task. |
| Researcher Affiliation | Academia | Shahira Abousamra1, Minh Hoai1, Dimitris Samaras1, Chao Chen2 1Department of Computer Science, Stony Brook University, USA 2Department of Biomedical Informatics, Stony Brook University, USA |
| Pseudocode | No | The paper describes the algorithm steps in narrative text and figures, but it does not include any formal pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | 1The code and a full version of this paper can be found at https://github.com/Topo XLab/Topo Count. |
| Open Datasets | Yes | We validate our method on popular crowd counting benchmarks including Shanghai Tech parts A and B (Zhang et al. 2016), UCF CC 50 (Idrees et al. 2013), UCF QNRF (Idrees et al. 2018), JHU++ (Sindagi, Yasarla, and Patel 2020), and NWPU Challenge (Wang et al. 2020c). |
| Dataset Splits | Yes | We use 50 50 pixels patches for Shanghai Tech and UCF CC 50, and 100 100 pixels patches for the larger scale datasets UCF QNRF, JHU++, and NWPU to account for the larger scale variation. An ablation study on the patch size selection is reported in the experiments. The persistence loss is applied on grid tiles to enforce topological consistency between corresponding prediction and ground truth tiles/patches. As data augmentation, coordinates of the top left corner of the grid are randomly perturbed. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions using a U-Net style architecture with a VGG-16 encoder and DICE loss, but it does not specify any software dependencies with version numbers (e.g., TensorFlow, PyTorch, or specific library versions). |
| Experiment Setup | Yes | We train our Topo Count with the dilated ground truth dot mask. The dilation is by default up to 7 pixels... We use 50 50 pixels patches for Shanghai Tech and UCF CC 50, and 100 100 pixels patches for the larger scale datasets UCF QNRF, JHU++, and NWPU... In the beginning of training we use DICE loss only (λ = 0). When the model starts to converge to reasonable likelihood maps, we add the persistence loss with λ = 1. |