A Novel Unified Architecture for Low-Shot Counting by Detection and Segmentation

Authors: Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, Matej Kristan

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments. Evaluation metrics and datasets. Standard datasets are used. The FSCD147 [19] is a detection-oriented extension of the FSC147 [24], which contains 6135 images of 147 object classes, split into 3659 training, 1286 validation, and 1190 test images. 4.1 Experimental Results. Table 1: Few-shot density-based methods (top part) and detection-based methods (bottom part) performances on the FSCD147 [19].
Researcher Affiliation Academia Jer Pelhan, Alan Lukežiˇc, Vitjan Zavrtanik, Matej Kristan Faculty of Computer and Information Science, University of Ljubljana jer.pelhan@fri.uni-lj.si
Pseudocode No The paper describes the method using detailed text and block diagrams but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes The code is available on Git Hub.
Open Datasets Yes Standard datasets are used. The FSCD147 [19] is a detection-oriented extension of the FSC147 [24], which contains 6135 images of 147 object classes, split into 3659 training, 1286 validation, and 1190 test images. The second dataset is FSCD-LVIS [19], derived from LVIS [8] and contains 377 categories.
Dataset Splits Yes The FSCD147 [19] is a detection-oriented extension of the FSC147 [24], which contains 6135 images of 147 object classes, split into 3659 training, 1286 validation, and 1190 test images.
Hardware Specification Yes The training is done on 2 A100s GPUs with standard scale augmentation [20; 4] and zero-padding images to 1024 1024 resolution.
Software Dependencies No The paper mentions software components like the SAM backbone and Adam W optimizer but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes Ge Co is pretrained with the classical loss [20] for initialization and is then trained for 200 epochs with the proposed dense detection loss (3) using a mini-batch size of 8, Adam W [16] optimizer, with initial learning rate set to 10 4, and weight decay of 10 4.