Thank you very much for your great work! Can you provide a version of the Oxford dataset that is the same as the clean Nordland dataset?
I'm sorry to bother you.

hi :)
Can you provide a pretrained model？
thanks

Hello @oravus ,

Thanks for your fantastic job providing a nice way to fuse information from multiple frames. Now I only use SeqNet (Conv+SAP+L2-Norm) without SeqMatch, trained with my own dataset for LIDAR-based place recognition. However, I found SeqNet did not work well once the seq_len is small (< 20) and was hard to train to get a strong model. Thus I guess that whether SeqNet without SeqMatch works well largely depends on the distribution of training set and test set, or it largely depends on the output module of the raw descriptor generation algorithm. Is this right? Have you met some dataset where SeqNet can not work well?

Best wishes!

Hello @oravus ,

Thanks for your fantastic job providing a nice way to fuse information from multiple frames. Now I only use SeqNet (Conv+SAP+L2-Norm) without SeqMatch, trained with my own dataset for LIDAR-based place recognition. However, I found SeqNet did not work well once the seq_len is small (< 20) and was hard to train to get a strong model. Thus I guess that whether SeqNet without SeqMatch works well largely depends on the distribution of training set and test set, or it largely depends on the output module of the raw descriptor generation algorithm. Is this right? Have you met some dataset where SeqNet can not work well?

Best wishes!

Hi, thanks for your remarkable work!
Could you please provide the original images dataset for training SeqNet? such as Nordland, Oxford, Brisbane, and MSLS? Thanks.

Hi! I am now working on your SeqNet project, and meeting some problems through it.
Currently I can run the codes and get recall results using nordland and oxford dataset. Now I am trying to use other datasets like hilti: https://hilti-challenge.com/dataset.html . I wonder how I can fit this dataset into your program.
Should I also generate .db AND .npy files as well? If my original data format is .jpg, what should be my process? the .npy files under /netvlad-pytorch file are descriptors generated by netVLAD, am I right?

Thank you for your patience and waiting for your kindly help!

Hi, I would like to replicate the results on Mapillary. Thank you for adding the code to generate the .npy files containing the descriptors for any dataset. Could you please suggest how to generate the .db files for MSLS and then replicate the results in your paper?
Thank your help!

Hi @oravus , thanks for your help and I have generated the needed db files using pitts250k dataset. Generally, I use the whole dataset to generate descriptors and saved both query and ref .npy files.
That is :
===> Loading dataset(s)
All Db descs: (254064, 4096)
All Qry descs: (24000, 4096)

Next, I use the first 10000 images as train sets, followed by 3000 images as validation and 3000 images as test sets. Also, I generated 3 .db files as train_mat_file, test_mat_file, val_mat_file.

Thereafter, I wrote the specification in get_datasets.py. The indexes are defined as nordland dataset format:
trainInds, testInds, valInds = np.arange(10000), np.arange(10000,13000), np.arange(13000,16000)

And I think I can test the dataset using your pretrained model, but things comes out that still errors occur regarding index.

//////////////////////////////////////////
Restored flags: ['--optim', 'SGD', '--lr', '0.0001', '--lrStep', '50', '--lrGamma', '0.5', '--weightDecay', '0.001', '--momentum', '0.9', '--seed', '123', '--runsPath', './data/runs', '--savePath', './data/runs/Jun03_15-22-44_l10_l10_w5_seqnetEnv/checkpoints', '--patience', '0', '--pooling', 'seqnet', '--w', '5', '--outDims', '4096', '--margin', '0.1']
Namespace(batchSize=16, cacheBatchSize=24, cachePath='./data/cache', cacheRefreshRate=0, ckpt='latest', dataset='pitts250k', descType='netvlad-pytorch', evalEvery=1, expName='0', extractOnly=False, lr=0.0001, lrGamma=0.5, lrStep=50.0, margin=0.1, mode='test', momentum=0.9, msls_trainCity='melbourne', msls_valCity='austin', nEpochs=200, nGPU=1, nocuda=False, numSamples2Project=-1, optim='SGD', outDims=4096, patience=0, pooling='seqnet', predictionsFile=None, resultsPath=None, resume='./data/runs/Jun03_15-22-44_l10_w5/', runsPath='./data/runs', savePath='./data/runs/Jun03_15-22-44_l10_l10_w5_seqnetEnv/checkpoints', seed=123, seqL=5, seqL_filterData=None, split='test', start_epoch=0, threads=8, w=5, weightDecay=0.001)
===> Loading dataset(s)
All Db descs: (254064, 4096)
All Qry descs: (24000, 4096)
===> Evaluating on test set
====> Query count: 800
===> Building model
=> loading checkpoint './data/runs/Jun03_15-22-44_l10_w5/checkpoints/checkpoint.pth.tar'
=> loaded checkpoint './data/runs/Jun03_15-22-44_l10_w5/checkpoints/checkpoint.pth.tar' (epoch 200)
===> Running evaluation step
====> Extracting Features
20%|████████████████████████████████████▉ | 49/249 [00:00<00:01, 127.41it/s]==> Batch (50/250)
39%|█████████████████████████████████████████████████████████████████████████▏ | 97/249 [00:00<00:01, 148.19it/s]==> Batch (100/250)
58%|████████████████████████████████████████████████████████████████████████████████████████████████████████████▉ | 145/249 [00:01<00:00, 154.13it/s]==> Batch (150/250)
78%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▉ | 193/249 [00:01<00:00, 156.80it/s]==> Batch (200/250)
97%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████▋ | 242/249 [00:01<00:00, 158.98it/s]==> Batch (250/250)
Average batch time: 0.006786982536315918 0.009104941585046229
torch.Size([3000, 4096]) torch.Size([3000, 4096])
====> Building faiss index
====> Calculating recall @ N
Using Localization Radius: 25
Traceback (most recent call last):
File "main.py", line 133, in
recallsOrDesc, dbEmb, qEmb, rAtL, preds = test(opt, model, encoder_dim, device, whole_test_set, writer, epoch, extract_noEval=opt.extractOnly)
File "/home/lx/lx/Seqnet_new/test.py", line 138, in test
rAtL.append(getRecallAtN(n_values, predictions, gtAtL))
File "/home/lx/lx/Seqnet_new/test.py", line 37, in getRecallAtN
if len(gt[qIx]) == 0:
IndexError: list index out of range

/////////////////////////////////////////
I am writing to ask :

1. how should the index should be defined? Could you use oxford dataset or nordland dataset to explain the details for me?
2. should npy files contain the same number of descriptors as I generate the dataset as 3 db files for train, test, and val?

THANKS FOR HELP!

Hi! Thanks for your work!
I have some doubts about the nordland image descriptors.

1. The nordland image descriptors are from netvlad. Whether the netvlad model was trained in the nordland?
2. Which one is a fixed distance sample between oxford-pnv and oxford-v1.0

Kind regards!

Hi! Thanks for your great work!
The argument, “seqL_filterData”, has always defaulted to None, What does it do and under what conditions is it used?
Kind regards!

Hello @oravus , thanks for your work. I was trying to run the method locally but I find that links provided for -

1. Dataset Download - https://cloudstor.aarnet.edu.au/plus/s/8L7loyTZjK0FsWT
2. Pre-trained models - https://cloudstor.aarnet.edu.au/plus/s/wx0zIGi3WBTtq5F are broken. Could you please provide the updated links? Thanks in advance.

I'm sorry to bother you.
I've noticed that the 'nGPU' parameter only appears once in the main function, not anywhere else.I would like to know how this parameter is used.
Best wishes to you!

Hi! Thanks for your great work on SeqNet.
The evaluation metric in SeqNet is only Recall@K, but the false positive instances is not considered.
Would you mind providing some implementation ideas in test step for evaluating the Precision@K?
Best wishes!

Hi! Thanks for your great work on SeqNet.
And I met a problem when I trained the sequential descriptors on the MSLS dataset. I just run the command: python main.py --mode train --pooling seqnet --dataset msls --msls_trainCity melbourne --msls_valCity austin --seqL 5 --w 3 --outDims 4096 --expName "msls_w3".
But ther is a problem: No such file or directory: './data/descData/netvlad-pytorch/msls_melbourne_databast.npy'.
Maybe I did not find any corresponding ***.npy files about the MSLS dataset. Would you mind providing the corresponging download link of the MSLS dataset.?
Kind regards!

Hi Dr. Garg
We recently came up with a new challenging dataset called https://mcdviral.github.io/

We'd like to invite you to try the SeqNet on these 3 campuses located at different latitudes.

if you have other IROS paper for me to review, I'm perfectly ok to do so

Regards
Shenghai Yuan

Hello @oravus ,

Thanks for your fantastic job providing a nice way to fuse information from multiple frames. Now I only use SeqNet (Conv+SAP+L2-Norm) without SeqMatch, trained with my own dataset for LIDAR-based place recognition. However, I found SeqNet did not work well once the seq_len is small (< 20) and was hard to train to get a strong model. Thus I guess that whether SeqNet without SeqMatch works well largely depends on the distribution of training set and test set, or it largely depends on the output module of the raw descriptor generation algorithm. Is this right? Have you met some dataset where SeqNet can not work well?

Best wishes!

@oravus Hi! Thanks for your kindly reply in issue #17! That do help me a lot!

In case my additional comments in issue #17 are not noticed I submit this new issue and copy my question here:

In the experiments on MSLS dataset (Fig.4 in the paper), Amman, Boston, San Francisco and Copenhagen are used for testing but only files related Amman are provided (imageNames .txt and seqBounds.txt), for the rest three cities should I use subsets like your descriptions above? In this case could you please provide these related files so that I could do some experiments and get the same results? As the principle to generate those subsets was not mentioned in the paper:).

Or it is sufficient to generate these files myself from MSLS dataset, which means, to use the whole set of original Boston, San Francisco and Copenhagen traverses.

Thank you again!

Hi! Thanks for your great work on SeqNet.
With your kindly help in issue #12 opened by myself, I have obtained the corresponding descriptor data for the Nordland, Oxford, and MSLS datasets, and achieved the same results as the paper. Then, I found that the Bribane dataset is not open-source, such that I cannot directly obtain the corresponding descriptor through the original dataset imsges. I have found your reply--"For the Brisbane dataset, we can only release the descriptor data." in issue #1 that the descritpor data can be released.
Would you mind providing the descriptor data of the Brisbane dataset for me? I just want to reproduce all the results in the paper and improve this remarkable work. I promise it will never be commercial but for research.
Kind regards!

Hi May I ask what the sequential descriptor really encodes?

As far as I understand, I can use LSTM to estimate both self-motion as well as motion stereo depth.

Does LSTM here only encodes the changes in locomotion? or does it also encode the overall 3D structure prior?

If it encodes either motion or map, I can use short-duration odometry as additional field input or something like LIDAR projected depth map to speed up the seqNet process to make it multi-modality seqNet?

I find that the number of images in original seq_info.csv and the number of images in seqBounds.txt in this code are not the same.

For example, in melbourne/query/seq_info.csv (original MSLS dataset) there are 88119 images but here in structFiles/seqBoundsFiles/msls_melbourne_query_seqBounds.txt only 4474! This is the same case for austin city, but for amman it's the same.

I wonder whether some of you know the reason about this difference...

Thanks a lot!!!

hi, can you provide a pretrained model on MSLS datasets？
thanks

Hi, thanks for sharing your work! I have two doubts reagrding the MSLS sequences.

1. In the paper, it is mentioned that for MSLS, the sequences were used as such. I checked the seqBound txt files and the sequence length seems to span hundreds of frames and is not fixed. However, it is said that the focus is on shorter sequences for VPR. Could you please explain this?
2. It is also mentioned in the paper that for all sequences, only those images are considered where a valid sequence can be 'centered'. However, looking at the seqBound txt files, it looks like the approach of taking the neighboring frames keeping the current frame as center is not used and the same frames are used for all frames in the sequence (e.g. 1 to 31 for all 31 frames). Could you please explain this as well?

Thanks in advance!