SeqNet: Learning Descriptors for Sequence-Based Hierarchical Place Recognition

[ArXiv+Supplementary] [IEEE Xplore RA-L 2021] [ICRA 2021 YouTube Video]


SeqNetVLAD vs PointNetVLAD: Image Sequence vs 3D Point Clouds for Day-Night Place Recognition

[ArXiv] [CVPR 2021 Workshop 3DVR]

Sequence-Based Hierarchical Visual Place Recognition.


Jan 18, 2022 : MSLS training setup included.

Jan 07, 2022 : Single Image Vanilla NetVLAD feature extraction enabled.

Oct 13, 2021 : Oxford & Brisbane Day-Night pretrained models download link.

Aug 03, 2021 : Added Oxford dataset files and a direct link to download the Nordland dataset.

Jun 23, 2021: CVPR 2021 Workshop 3DVR paper, "SeqNetVLAD vs PointNetVLAD", now available on arXiv.



conda create -n seqnet numpy pytorch=1.8.0 torchvision tqdm scikit-learn faiss tensorboardx h5py -c pytorch -c conda-forge


Run bash to download single image NetVLAD descriptors (3.4 GB) for the Nordland-clean dataset [a] and the Oxford dataset (0.3 GB), and Nordland-trained model files (1.5 GB) [b]. Other pre-trained models for Oxford and Brisbane Day-Night can be downloaded from here.



To train sequential descriptors through SeqNet on the Nordland dataset:

python --mode train --pooling seqnet --dataset nordland-sw --seqL 10 --w 5 --outDims 4096 --expName "w5"

or the Oxford dataset (set --dataset oxford-pnv for pointnetvlad-like data split as described in the CVPR 2021 Workshop paper):

python --mode train --pooling seqnet --dataset oxford-v1.0 --seqL 5 --w 3 --outDims 4096 --expName "w3"

or the MSLS dataset (specifying --msls_trainCity and --msls_valCity as default values):

python --mode train --pooling seqnet --dataset msls --msls_trainCity melbourne --msls_valCity austin --seqL 5 --w 3 --outDims 4096 --expName "msls_w3"

To train transformed single descriptors through SeqNet:

python --mode train --pooling seqnet --dataset nordland-sw --seqL 1 --w 1 --outDims 4096 --expName "w1"


On the Nordland dataset:

python --mode test --pooling seqnet --dataset nordland-sf --seqL 5 --split test --resume ./data/runs/Jun03_15-22-44_l10_w5/ 

On the MSLS dataset (can change --msls_valCity to melbourne or austin too):

python --mode test --pooling seqnet --dataset msls --msls_valCity amman --seqL 5 --split test --resume ./data/runs/<modelName>/

The above will reproduce results for SeqNet (S5) as per Supp. Table III on Page 10.

[Expand this] To obtain other results from the same table in the paper, expand this.
# Raw Single (NetVLAD) Descriptor
python --mode test --pooling single --dataset nordland-sf --seqL 1 --split test

# SeqNet (S1)
python --mode test --pooling seqnet --dataset nordland-sf --seqL 1 --split test --resume ./data/runs/Jun03_15-07-46_l1_w1/

# Raw + Smoothing
python --mode test --pooling smooth --dataset nordland-sf --seqL 5 --split test

# Raw + Delta
python --mode test --pooling delta --dataset nordland-sf --seqL 5 --split test

# Raw + SeqMatch
python --mode test --pooling single+seqmatch --dataset nordland-sf --seqL 5 --split test

# SeqNet (S1) + SeqMatch
python --mode test --pooling s1+seqmatch --dataset nordland-sf --seqL 5 --split test --resume ./data/runs/Jun03_15-07-46_l1_w1/

# HVPR (S5 to S1)
# Run S5 first and save its predictions by specifying `resultsPath`
python --mode test --pooling seqnet --dataset nordland-sf --seqL 5 --split test --resume ./data/runs/Jun03_15-22-44_l10_w5/ --resultsPath ./data/results/
# Now run S1 + SeqMatch using results from above (the timestamp of `predictionsFile` would be different in your case)
python --mode test --pooling s1+seqmatch --dataset nordland-sf --seqL 5 --split test --resume ./data/runs/Jun03_15-07-46_l1_w1/ --predictionsFile ./data/results/Jun03_16-07-36_l5_0.npz

Single Image Vanilla NetVLAD Extraction

[Expand this] To obtain the single image vanilla NetVLAD descriptors (i.e. the provided precomputed .npy descriptors)
# Setup Patch-NetVLAD submodule from the seqNet repo:
cd seqNet 
git submodule update --init

# Download NetVLAD+PCA model
cd thirdparty/Patch-NetVLAD/patchnetvlad/pretrained_models
wget -O pitts_orig_WPCA4096.pth.tar

# Compute global descriptors
cd ../../../Patch-NetVLAD/
python --config_path patchnetvlad/configs/seqnet.ini --dataset_file_path ../../structFiles/imageNamesFiles/oxford_2014-12-16-18-44-24_imagenames_subsampled-2m.txt --dataset_root_dir <PATH_TO_OXFORD_IMAGE_DIR> --output_features_fullpath ../../data/descData/netvlad-pytorch/oxford_2014-12-16-18-44-24_stereo_left.npy

# example for MSLS (replace 'database' with 'query' and use different city names to compute all)
python --config_path patchnetvlad/configs/seqnet.ini --dataset_file_path ../../structFiles/imageNamesFiles/msls_melbourne_database_imageNames.txt --dataset_root_dir <PATH_TO_Mapillary_Street_Level_Sequences> --output_features_fullpath ../../data/descData/netvlad-pytorch/msls_melbourne_database.npy


The code in this repository is based on Nanne/pytorch-NetVlad. Thanks to Tobias Fischer for his contributions to this code during the development of our project QVPR/Patch-NetVLAD.


  title={SeqNet: Learning Descriptors for Sequence-based Hierarchical Place Recognition},
  author={Garg, Sourav and Milford, Michael},
  journal={IEEE Robotics and Automation Letters},

  title={SeqNetVLAD vs PointNetVLAD: Image Sequence vs 3D Point Clouds for Day-Night Place Recognition},
  author={Garg, Sourav and Milford, Michael},
  howpublished={CVPR 2021 Workshop on 3D Vision and Robotics (3DVR)},

Other Related Projects

SeqMatchNet (2021); Patch-NetVLAD (2021); Delta Descriptors (2020); CoarseHash (2020); seq2single (2019); LoST (2018)

[a] This is the clean version of the dataset that excludes images from the tunnels and red lights and can be downloaded from here.

[b] These will automatically save to ./data/, you can modify this path in and to specify your workdir.

CRICOS No. 00213J