This repository corresponds to the work entitled 'Creation of Topological Maps by Visual Appearance by Cutting Graphs. Applications to the Recognition of Previously Visited Places' as a Bachelor's Thesis.
Authors: Ángel Montejo Quesada and Javier Gonzalez-Jimenez
This repository implements a method for creating topological maps for Visual Place Recognition based on appearance. It is founded on the creation and segmentation of an Appearance Graph.
-AppearanceGraph_Construction: This code implements the computation of the co-visibility matrix and similarity matrix from a given Appearance Map, which leads to the construction of the adjacency matrix for the Appearance Graph.
-AppearanceGraph_NormalizedCut: This code implements the recursive bipartition of the Appearance Graph using the Normalized Cut method, resulting in a topology of places for the original Appearance Map.
-Covisibility_evaluation: This code is used to evaluate the creation of the Appearance Graph, focusing on the computation of the co-visibility matrix and studying the involved parameter. For this evaluation, a controlled dataset is needed where the actual co-visibility images are known. This code concludes the optimum value for the parameter in the co-visibility measure.
-VisualPlaceRecognition_EvaluationTools: This code implements a Visual Place Recognition system, along with different tools to evaluate the Normalized Cut of the Appearance Graph.
-CutGraph_evaluation: This code uses the tools implemented in 'VisualPlaceRecognition_EvaluationTools' to study different aspects of the Normalized Cut of the Appearance Graph. It concludes the optimum values of the involved parameters, along with the optimal configuration for this method.
.../outputs This folder stores the results of the Appearance Graph construction and segmentation.
.../Evaluation In this folder, the results of the 'CutGraph_evaluation' are stored for later comparison.
.../dataset/map This folder corresponds to the dataset representing the original Appearance Map.
.../dataset/map/Poses This folder contains the .json file with the pose and image names information, among other details.
.../dataset/map/Descriptors/netVLAD This folder contains the .h5py file with the netVLAD descriptors for the images in the Appearance Map.
.../dataset/map/Images This folder contains the raw images for the given Appearance Map, used for the co-visibility computation.
.../dataset/query This folder corresponds to the query dataset used for the Visual Place Recognition task. The directory structure is the same as the '.../dataset/map' directory.
This software employs built-in libs (see requeriments.txt), and has been tested with Python>=3.7.15 on conda 22.11.1
The geometry2.py script is inspired in ProbFiltersVPR.