Particular object retrieval with integral max-pooling of CNN activations
Authors: Giorgos Tolias, Ronan Sicre, Hervé Jégou
ICLR 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This paper significantly improves existing CNN-based recognition pipeline: We report for the first time results competing with traditional methods on the challenging Oxford5k and Paris6k datasets. (abstract) and This section presents the results of our compact representation for image retrieval, evaluate the localization accuracy AML, and finally employ it for retrieval re-ranking. Experimental setup. We evaluate the proposed methods on Oxford Buildings dataset (Philbin et al., 2007) and Paris dataset (Philbin et al., 2008) (Section 8, Experimental setup) |
| Researcher Affiliation | Collaboration | Giorgos Tolias Center for Machine Perception FEE CTU Prague Ronan Sicre Irisa Rennes Herv e J egou Facebook AI Research |
| Pseudocode | No | The paper describes methods in prose but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statement or link regarding the availability of its source code. |
| Open Datasets | Yes | We evaluate the proposed methods on Oxford Buildings dataset (Philbin et al., 2007) and Paris dataset (Philbin et al., 2008), which are composed of 5063 and 6412 images, respectively. We refer to these datasets as Oxford5k and Paris6k. |
| Dataset Splits | No | The paper mentions 'PCA is learned on Paris6k when testing on Oxford5k and vice versa' and 'We follow the standard protocol', but it does not provide specific training, validation, or test dataset splits needed for reproduction beyond the general benchmark protocol. |
| Hardware Specification | No | The paper mentions using 'Alex Net' and 'VGG16' networks and 'Mat Conv Net' for feature extraction, but it does not specify any hardware details like CPU or GPU models, or memory. |
| Software Dependencies | No | Mat Conv Net (Vedaldi & Lenc, 2014) is used to extract the features. (Section 8, Experimental setup). The paper mentions software by name but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | We set α equal to 10 in all of our experiments. (Section 5, Approximate integral max-pooling) and We finally set s = 1.1 and t = 3 for re-ranking usage. (Section 8, Localization accuracy) and We evaluate different input image resolutions and observe that the original image size (1024) provides higher performance. (Section 8, Retrieval and re-ranking) |