R2D2: Reliable and Repeatable Detector and Descriptor

Authors: Jerome Revaud, Cesar De Souza, Martin Humenberger, Philippe Weinzaepfel

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiment results show that our elegant formulation of joint detector and descriptor selects keypoints which are both repeatable and reliable, leading to state-of-the-art results on the HPatches and Aachen datasets.
Researcher Affiliation Industry Jerome Revaud Philippe Weinzaepfel César De Souza Martin Humenberger NAVER LABS Europe firstname.lastname@naverlabs.com
Pseudocode No No structured pseudocode or algorithm blocks are explicitly provided or labeled in the paper.
Open Source Code Yes Our code and models are available at https://github.com/naver/r2d2.
Open Datasets Yes We use three sources of data to train our method: (a) distractors from a retrieval dataset [37] (i.e., random web images), from which we build synthetic image pairs by applying random transformations (homography and color jittering), (b) images from the Aachen dataset [45, 47], using the same strategy to build synthetic pairs, and (c) pairs of nearby views from the Aachen dataset where we obtain a pseudo ground-truth using optical flow (see below). All sources are represented approximately equally (about 4000 images each) and we study their importance in Section 4.4. Note that we do not use any image from the HPatches evaluation dataset [2] during training.
Dataset Splits No The paper does not explicitly provide specific training/validation/test dataset splits (e.g., percentages or exact counts for validation). It mentions training data and test data (HPatches, Aachen) but no explicit validation set split or methodology for it within the main text.
Hardware Specification Yes In practice, processing a 1M pixel image on a Tesla P100-SXM2 GPU takes about 0.5s to extract keypoints at a single scale (full image) and 1s for all scales.
Software Dependencies No The paper does not provide specific version numbers for software dependencies (e.g., libraries like PyTorch, TensorFlow, or specific Python versions) used in the experiment.
Experiment Setup Yes We optimize the network using Adam for 25 epochs with a fixed learning rate of 0.0001, weight decay of 0.0005 and a batch size of 8 pairs of images cropped to 192 192.