Delta Descriptors


Update 2021-Jun-02: A pytorch-based (GPU/CPU) implementation of Delta Descriptors is now available with our latest work SeqNet.

Delta Descriptors

Source code for the paper - "Delta Descriptors: Change-Based Place Representation for Robust Visual Localization", published in IEEE Robotics and Automation Letters (RA-L) 2020 and to be presented at IROS 2020. [arXiv] [IEEE Xplore][YouTube]

We propose Delta Descriptor, defined as a high-dimensional signed vector of change measured across the places observed along a route. Using a difference-based description, places can be effectively recognized despite significant appearance variations. Schematic of the proposed approach Images on the left are from the Oxford Robotcar dataset.



See requirements.txt, generated using pipreqs==0.4.10 and python3.5.6


Download this Repository and the Nordland dataset (part)

The dataset used in our paper is available here (or use commands as below). Note that the download only comprises a small part (~1 GB) of the original Nordland videos released here. These videos were first used for visual place recognition in this paper.

git clone
cd DeltaDescriptors/
mkdir data/
cd data/

The zip contains two folders: summer and winter, where each one of them comprises 1750 images which were used for experiments conducted in our paper.

Describe and Match

Delta Descriptors are defined on top of global image descriptors, for example, NetVLAD (Update 05 Sep 2020: see our python wrapper). Given such descriptors, compute Delta Descriptors and match across two traverses as below:

python src/ --genDesc --genMatch -l 16 -d delta -ip1 <full_path_of_desc.npy> -ip2 <full_path_of_query_desc.npy>

The input descriptor data is assumed to be a 2D tensor of shape [numImages,numDescDims]. The computed descriptors are stored in .npy format and the match results are stored in .npz format comprising a dict of two arrays: matchInds (matched reference index per query image) and matchDists (corresponding distance value). By default, output is stored in the ./out folder but can also be specified via --outPath argument. To see all the options, use:

python src/ --help

The options --genDesc and --genMatch can be used in isolation or together, see example usage below.

Describe only

In order to compute only the descriptors for a single traverse, use:

python src/ --genDesc -l 16 -d delta -ip1 <full_path_of_desc.npy>

Match only

For only computing matches, given the descriptors (Delta or some other), use:

python src/ --genMatch -ip1 <full_path_of_desc.npy> -ip2 <full_path_of_query_desc.npy>

Evaluate only

python src/ --eval -mop <full_path_of_match_output.npz>

or evaluate directly with --genMatch (and possibly --genDesc) flag:

python src/ --eval --genMatch -ip1 <full_path_of_desc.npy> -ip2 <full_path_of_query_desc.npy>

Currently, only Nordland dataset-style (1-to-1 frame correspondence) evaluation is supported, GPS/INS coordinates-based evaluation, for example, for Oxford Robotcar dataset to be added soon. Evalution code can be used to generate PR curves and the code in its current form prints Precision @ 100% Recall for localization radius of 1, 5, 10 and 20 (frames).


If you find this code or our work useful, cite it as below:

  title={Delta Descriptors: Change-Based Place Representation for Robust Visual Localization},
  author={Garg, Sourav and Harwood, Ben and Anand, Gaurangi and Milford, Michael},
  journal={IEEE Robotics and Automation Letters},


The code is released under MIT License.

Related Projects

SeqNet (2021).

CoarseHash (2020)

seq2single (2019)

LoST (2018)

CRICOS No. 00213J