PBECount: Prompt-Before-Extract Paradigm for Class-Agnostic Counting

Authors: Canchen Yang, Tianyu Geng, Jian Peng, Chun Xu

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results on the FSC-147 and CARPK datasets demonstrate that the proposed PBECount can identify whether unknown class objects are similar to exemplars and outperform the state-of-the-art CAC methods in terms of accuracy and generalization. The verification results indicated that the proposed PBECount can achieve excellent counting performance on the FSC-147 dataset (Ranjan et al. 2021), with MAE values of 8.88 and 7.71 on the validation and test sets, respectively. Experiments Implementation Details To demonstrate the superiority of the PBECount model, as well as the ability of the proposed prompt-before-extract paradigm to assess the exemplar-object similarity, this study conducted experiments on the FSC-147 dataset (Ranjan et al. 2021), and adopted the mean absolute error (MAE) and root mean squared error (RMSE) as evaluation metrics. Comparison with State-of-the-Art Methods Quantitative Results The counting performance of the proposed PBECount was compared with other CAC methods on the FSC-147 dataset, as presented in Table 1. Ablation Study The ablation experiments were conducted on the FSC-147 dataset, analyzing the impacts of individual architectural components and peak-aware MSE loss on the counting performance of the PBECount.
Researcher Affiliation Academia 1 College of Computer Science, Sichuan University, Chengdu, 610065, China 2 Information Construction and Management Office, Sichuan University, 610065, China EMAIL, EMAIL
Pseudocode No The paper describes methods and processes through text and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain an unambiguous statement about releasing code or a direct link to a source-code repository for the methodology described.
Open Datasets Yes The experimental results on the FSC-147 and CARPK datasets... The verification results indicated that the proposed PBECount can achieve excellent counting performance on the FSC-147 dataset (Ranjan et al. 2021)... Moreover, it demonstrates superior cross-dataset generalization performance on the CARPK dataset (Hsieh, Lin, and Hsu 2017).
Dataset Splits Yes The verification results indicated that the proposed PBECount can achieve excellent counting performance on the FSC-147 dataset (Ranjan et al. 2021), with MAE values of 8.88 and 7.71 on the validation and test sets, respectively. ...To further investigate the counting performance of the proposed PBECount for images of different density levels, this study partitioned the validation and test sets of the FSC-147 dataset into eight subsets based on the number of objects in each image. ...Next, 66 images were selected from the validation set, and 33 images were selected from the test set of the FSC-147 dataset. These images all contained two or more object categories with a number larger than three, forming the FSC-147 Mul dataset. ...For counting purposes, three objects were randomly sampled from each image in the CARPK test set as exemplars.
Hardware Specification Yes The training and testing processes of the proposed model were conducted on an NVIDIA Ge Force RTX 4090D GPU with a batch size of one for approximately 25 hours.
Software Dependencies No The paper mentions using the AdamW optimizer and specific loss functions (BCE with logits loss, peak-aware MSE loss), but does not provide specific version numbers for any programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes The proposed model was trained for 250 epochs using the Adam W optimizer with a weight decay of 10 4. During the first 200 training epochs, the cosine annealing strategy was used to adjust the learning rate, starting from an initial value of 10 4 and gradually decreasing it to 10 5. Subsequently, the model weight that achieved an optimal counting performance was fine-tuned using an additional fixed learning rate of 10 5 for 50 epochs. The training and testing processes of the proposed model were conducted on an NVIDIA Ge Force RTX 4090D GPU with a batch size of one for approximately 25 hours.