I think your implementation of best_pos_distance (which is used for every loss in your code) should use reduce_max instead of reduce_min (following the original hard triplet loss paper). I re-implemented your code in PyTorch, and with max pos_distance i got an average top 1% recall on oxford baseline (no refine) of 84%

Hi, i was running the code using a computer with 16GB memory and NVIDIA GeForce GTX2070. However, the memory was filled when I run the generate_training_tuples_baselineand.py & evaluate.py. The process was killed by the system. It seems that the size of the memory is not enough. Therefore, I would like to know how large the RAM did you use?

@mikacuy Hi，mika. First of all, thanks for your great work. I'm new to place recognition tasks and trying to retrain your network as a tutorial. I used to retrain some other networks and the losses go almost down directly while the loss defined in your network keeps going up and down. Here is a screenshot which I log out the batch losses. So does it go the right way as you did before? Tell me if I miss any details. Thanks

Claude Cui

Hi! Thanks for your great work and code release. Recently I am also doing some research on 3d-relocalization. After reading your paper and part of code, I am trying to convert the kitti data to the format which this code can run. And I wonder how strong is the generalization ability of this model? Will the recall decrease much? Have you tested before? Great thanks!!

Hi, thanks for your great job! But I am trying to visualize your submaps, and the following picture is from one of the .bin file However, the max value is up to 10^38. I have no idea what is the visualization result represent for. Is there something wrong with the submap dataset? Or I miss something?

Hi, thanks for your code and dataset. Could you please tell me the difference between the folder called pointcloud_20m_10overlap and the folder called pointcloud_20m?

Hi, @mikacuy may I ask how to solve this problem? It seems there is no file named model.ckpt. I only found model_baseline.ckpt.index and model_refine.ckpt.index from the downloaded pre-trained model.

2019-12-21 18:46:13.067473: W tensorflow/core/framework/op_kernel.cc:1192] Not found: Unsuccessful TensorSliceReader constructor: Failed to find any matching files for log/model.ckpt

Thanks.

Hi, thanks for your great work! I am trying to use the script generate_submaps.m to recreate the benchmark dataset for the Oxford Robotcar dataset. However, I seem to get different pointclouds (as described by the timestamps) than the ones found in the uploaded benchmarks folder on google drive.

For example, for the 2015-08-17-10-26-38 folder, the 1439803802060497.bin pointcloud cannot be found after generating the submaps, with the closest generated timestamps being: 1439803802982745.bin and 1439803801699224.bin.

There is some overlap between the generated and posted pointclouds for the same 2015-08-17-10-26-38 folder, such as: 1439803985022188.bin and 1439803989723175.bin.

I have not modified the uploaded script except for path changes, and am using the default values for the following parameters:

submap_cover_distance=20.
laser_reading_distance=0.025;
laser_reading_angle=30;
dist_start_next_frame=10.0;

Are there any parameter changes that need to be made to get the generated timestamps & pointclouds to correspond? Thank you!

Hi! Thanks for sharing your code! I couldn't find the source code of removing ground plane (e.g. pcfitplane). Could you give me some guidance? Is there something wrong? Thanks! Looking forward to your reply!

Hi, thanks for your code. I have a question regarding the datasets you are using to generate the training tuples. In the file generate_training_tuples_baseline.py, line 21, index_list=range(len(all_folders)-1) , it seems that you are using all runs except for the last sequence. Any reason for this? Or did I miss something? Thanks for your help in advance!

Thanks for your wonderful work, but i find in the file "train_pointnetvlad", the global variable "HARD_NEGATIVES={}" is never updated which means it is useless, is there something wrong here?

Dear Author,

It seems like PointNetVLAD couldn't not perform well for reverse direction loops in my experiments although there is T-Net... Have you tried to recognize reverse direction localization with PointNetVLAD? I'm not sure if my experiments went wrong.

Thanks.

Cannot find the following folder and file:

• Test/Inference submaps found in the folder "pointcloud_20m/" and its corresponding csv file is "pointcloud_locations_20m.csv"

Hi, thanks for your code! I have a question about training and test data split in generate_training_tuples_baseline.py. Line 30, you set p1 = [5735712.768124,620084.402381], p2 = [5735611.299219,620540.270327], p3 = [5735237.358209,620543.094379], and p4 = [5734749.303802,619932.693364]. I think they are the mean value of "northing" and "easting". I just want to know how to get the four different mean values. And why you choose four mean values not five or more. Thanks for your help in advance!

I have read your paper about PointNetVLAD and run the code you provide on the Github. It is really a great work.

Now, I would like to use PointNetVLAD for place localization (or point cloud retrieval) using the global descriptor. In fact, I see a lot of similar work you have done in your paper. For example, the work offigure 4. You used the query point clouds to retrieve the corresponding point clouds.

Now, I would like to perform the similar task as you did in figure 4. However, the code provided on the github is more likely to evaluate the performance of the network.

I would like to know what code you used to realize the above function (figure 4 in paper) so that it can get a retrieved point cloud with a input one. Otherwise, how can I use the existing code on github to fulfill this task?

In fact, I have sent this message to your email [email protected]. It seems that the address has expired. You may update it if you do not mind.

Thanks.