Active Image Indexing
Authors: Pierre Fernandez, Matthijs Douze, Herve Jegou, Teddy Furon
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that the retrieval and copy detection of activated images is significantly improved. For instance, activation improves by +40% the Recall1@1 on various image transformations, and for several popular indexing structures based on product quantization and locality sensitivity hashing. |
| Researcher Affiliation | Collaboration | Pierre Fernandez1,2, , Matthijs Douze1, Herv e J egou1, Teddy Furon2, 1Meta AI, FAIR 2Centre Inria de l Universit e de Rennes Correspondence: pfz@meta.com. |
| Pseudocode | Yes | Algorithm 1 Active indexing for IVF-PQ |
| Open Source Code | Yes | The implementation is available at github.com/facebookresearch/active indexing. |
| Open Datasets | Yes | The dataset DISC 2021 was designed for the Image Similarity Challenge (Douze et al., 2021) and can be downloaded in the dataset webpage: https://ai.facebook.com/datasets/disc21-dataset/. |
| Dataset Splits | No | The paper describes a training set for the index (1M images) and a query set for evaluation (50k images), but it does not explicitly mention a separate validation set for model tuning or early stopping. |
| Hardware Specification | No | The paper mentions running experiments 'with a 32GB GPU' but does not specify the exact model or type of the GPU, nor any other hardware components like CPU or memory. |
| Software Dependencies | No | The paper mentions that the implementation builds upon 'Pytorch (Paszke et al., 2019) and FAISS (Johnson et al., 2019) libraries' but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | Optimization (5) is done with the Adam optimizer (Kingma & Ba, 2015), the learning rate is set to 1, the number of iterations to N = 10 and the regularization to λ = 1. In (8), the distortion scaling is set to α = 3 (leading to an average PNSR around 43 d B). |