OmniCount: Multi-label Object Counting with Semantic-Geometric Priors
Authors: Anindya Mondal, Sauradip Nag, Xiatian Zhu, Anjan Dutta
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our comprehensive evaluation in Omni Count-191, alongside other leading benchmarks, demonstrates Omni Count s exceptional performance, significantly outpacing existing solutions. |
| Researcher Affiliation | Academia | Anindya Mondal1*, Sauradip Nag2*, Xiatian Zhu1, Anjan Dutta1 1University of Surrey 2Simon Fraser University EMAIL, EMAIL |
| Pseudocode | No | The paper describes the Omni Count pipeline and its components (Semantic Estimation, Geometric Estimation, Object Recovery, Reference point guided counting) using figures, equations, and descriptive text. However, it does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | The paper provides a web link: "https://mondalanindya.github.io/Omni Count". This is a project webpage and not a direct link to a source-code repository, nor does the paper contain an explicit statement about releasing the code for the methodology described. |
| Open Datasets | Yes | To evaluate Omni Count, we created the Omni Count-191 benchmark, a firstof-its-kind dataset with multi-label object counts, including points, bounding boxes, and VQA annotations. ... To address this gap, we create the Omni Count-191 benchmark. It includes 302,300 object instances across 191 categories in 30,230 images, featuring multiple categories per image and detailed annotations such as counts, points, and bounding boxes for each object (Fig. 4). ... We also benchmark on the PASCAL VOC dataset (Everingham et al. 2009), which contains 9963 images across 20 classes, with 4952 for testing. For singleclass counting, we use the test sets from FSC-147 (Ranjan et al. 2021) (1190 images, 29 categories) and CARPK (Hsieh, Lin, and Hsu 2017) (1014 test images). |
| Dataset Splits | Yes | To ensure diversity, the dataset is split into training and testing sets, with no overlap in object categories 118 categories for training and 73 for testing, corresponding to a 60%-40% split. This results in 26,978 images for training and 3,252 for testing. |
| Hardware Specification | No | The paper does not explicitly describe the hardware used to run its experiments. No specific GPU models, CPU models, or other hardware specifications are mentioned. |
| Software Dependencies | No | The paper mentions several models and frameworks used, such as Side Adapter Network (SAN), Marigold, Segment Anything Model (SAM), Grounding-DINO, and CLIPSeg. However, it does not provide specific version numbers for these software components or any underlying programming languages or libraries, which would be necessary for reproducibility. |
| Experiment Setup | No | The paper describes the methodology and evaluation process. While it discusses the structure of Omni Count and its components, it does not provide specific experimental setup details such as hyperparameter values (e.g., learning rate, batch size, number of epochs), optimizer settings, or other system-level training configurations in the main text. |