Vocabulary-free Image Classification

Authors: Alessandro Conti, Enrico Fini, Massimiliano Mancini, Paolo Rota, Yiming Wang, Elisa Ricci

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on benchmark datasets validate that Ca SED outperforms other complex vision-language frameworks, while being efficient with much fewer parameters, paving the way for future research in this direction1. We experiment on several datasets, considering both coarse(e.g. Caltech-101 [14], UCF101 [55]) and fine-grained (e.g. FGVC-Aircraft [40], Flowers-102 [43]) classification tasks.
Researcher Affiliation Collaboration Alessandro Conti1 Enrico Fini1 Massimiliano Mancini1 Paolo Rota1 Yiming Wang2 Elisa Ricci1,2 1University of Trento 2Fondazione Bruno Kessler (FBK)
Pseudocode No No. The paper describes the method steps in narrative form and with equations but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes Code and demo is available at https://github.com/altndrr/vic
Open Datasets Yes Datasets. We follow existing works [53, 66] and use ten datasets that feature both coarse-grained and fine-grained classification in different domains: Caltech-101 (C101) [14], DTD [7], Euro SAT (ESAT) [21], FGVC-Aircraft (Airc.) [40], Flowers-102 (Flwr) [43], Food-101 (Food) [4], Oxford Pets (Pets), Stanford Cars (Cars) [29], SUN397 (SUN) [61], and UCF101 (UCF) [55]. Additionally, we used Image Net [10] for hyperparameters tuning. As database, we use a subset of PMD [54], containing five of its largest datasets: Conceptual Captions (CC3M) [52], Conceptual Captions 12M (CC12M) [5], Wikipedia Image Text (WIT) [56], Redcaps [12], and a subset of [57] used for PMD (YFCC100M*).
Dataset Splits No No. The paper mentions using ImageNet for hyperparameter tuning but does not provide specific training/validation/test split percentages, sample counts, or citations to predefined splits for the main datasets (Caltech-101, DTD, etc.) used for evaluation.
Hardware Specification Yes Implementation details. Our experiments were conducted using NVIDIA A6000 GPUs with mixedbit precision.
Software Dependencies No No. The paper mentions the use of CLIP and the NLP library flair (https://github.com/flair NLP/flair) but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We tuned the α hyperparameter of Eq. (6) and the number of retrieved captions K of our method on the Image Net dataset, finding that α = 0.7 and K = 10 led to the best results. We use these values across all experiments.