Sequentially Generated Instance-Dependent Image Representations for Classification

Authors: Ludovic Denoyer; Matthieu Cord; Patrick Gallinari; Nicolas Thome; Gabriel Dulac-Arnold

ICLR 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the system s abilities in a series of image-based exploration and classification tasks that highlight its learned exploration and inference abilities. (010-012) and We present an experimental evaluation of this method on three different classical datasets and propose a qualitative study explaining the behaviour of this model. (265-268)
Researcher Affiliation Academia LIP6, UPMC Sorbonne University Paris, France (002-003)
Pseudocode Yes Algorithm 1 Inference Algorithm (385) and Algorithm 2 Complete Learning algorithm (440) and Algorithm 3 Classification Policy Learning Algorithm (495) and Algorithm 4 Exploration Sub-policy πγk Learn. Algorithm (550)
Open Source Code No The paper does not provide any explicit statement about open-sourcing its code or a link to a code repository.
Open Datasets Yes We evaluate the proposed method on two challenging image databases corresponding to different tasks: fine-grained image classification (People Playing Musical Instruments dataset5) (Yao & Fei-Fei, 2010) and scene recognition (15-scenes dataset) (Lazebnik et al., 2006). (615-619) and http://ai.stanford.edu/ bangpeng/ppmi.html (620)
Dataset Splits Yes We randomly sample training/testing images on 5 splits of the data, and the final performance corresponds to the average performance obtained on the 5 runs. (677-680)
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models used for experiments.
Software Dependencies No The paper mentions using 'VLFEAT library' but does not specify its version number or any other software dependencies with version details.
Experiment Setup Yes These local features are computed at a single scale (s = 16 pixels) and with a constant step size d. For 15-scenes we use d = 8 pixels while d = 4 pixels for PPMI. (623-625) and We run a K-Means algorithm by randomly sampling about 1 million descriptors in each database to produce a dictionnary of M = 200 codeword elements. (626-629) and We have chosen to create 10 sequences of regions for each training images, resulting in a training set of size 10 ℓfor each classifier fθ and πγt learned by our method. (643-646)