Test other datasets with models

[whyNotwh]

Could you please provide some more specific guidance so that I can use your method to test my dataset?Before using the method, in addition to the features of the dataset, are there any requirements for the format of the dataset and the documents required or anyother preparation?(such as dataset.csv and what are the content requirements under the relevant PATHs mentioned in setup.sh)

[mingu6]

Hi, thanks for taking an interest in the work. A few comments:

• dataset.csv is a file that is associated with the RobotCar dataset scraper https://github.com/mttgdd/RobotCarDataset-Scraper
• Dataset requirements currently align with the RobotCar dataset, see rtk.csv files for ground truth downloadable from https://robotcar-dataset.robots.ox.ac.uk/datasets/ as well as vo.csv files for odometry associated with individual traverses
• the paths in setup.py are used for intermediate data processing for RobotCar. Raw data downloaded from the website needs to be debayerised and RTK ground truth + vo need to be adjusted so we have pose estimates/relative pose estimates at image timestamps.
• Image descriptors are assumed to be stored in a single numpy array on disk named after the associated method (e.g. netvlad.npy). Rows are associated to images and sorted by timestamp value.

We will be releasing a new repo which will supersede this one (and be much neater) in the coming weeks with a complete data processing pipeline. To use this work for your own datasets, I would suggest running through a RobotCar query/reference pair to understand the data format for odom/image descriptors for the models in the models/ directory. Then you can implement your datasets as you see fit and use the data structures in models/ to run our methods and baselines.

[whyNotwh]

Thank you very much for your answer！I still have a few questions to ask you~

• Is my description of descriptor correct? For sequence retrieval, n_loc represents the number of querying images, t represents the sequence length, and dim=4096 represents the length of descriptor, so my format should be [n_loc, t, dim]. And I can set the size of t as 5, same thing with DB [n_db_loc, t, dim], right?If it's a single frame retrieval T is going to be one, right?
• Is threshold convergence detection used in Topological methods?
• And how should I understand your method of calculating the similarity of observed distances

  np.sqrt(2 - 2 * np.dot(self.map_descriptors, descriptor))

[mingu6]

Sorry for the late reply. To answer your questions:

• First point, correct, that is the right format.
• Single frame retrieval is 1
• Threshold detection is used in the Topological method
• For DB, do you mean the reference map? If so that'd just be [n_db, dim] in our case, we assume a single contiguous map instead of sequences for our methods. For say, the SeqSLAM comparison we do compare sequences to sequences where your interpretation would be correct [n_db, t, dim].
• So that formula you shared is computing the Euclidean descriptor between image descriptors. We can use this formula (faster to use np.dot instead of np.linalg.norm) since the descriptors are l2 normalised.

I will be tiding this up in the coming weeks with documentation, thanks for this!