kxhit / sg_pr Goto Github PK
View Code? Open in Web Editor NEWImplementation of IROS20 paper - "Semantic Graph Based Place Recognition for 3D Point Clouds"
License: MIT License
Implementation of IROS20 paper - "Semantic Graph Based Place Recognition for 3D Point Clouds"
License: MIT License
Hi!I want to know the time required for the network training, and is epochs 500?
Hello, I want to know how to run your code on the kitti dataset (referring to SG_PR). I hope to give more detailed steps in the readme document. I look forward to your answer, thank you.
I test pair_file: ["/SG_PR_DATA/graphs_sk/05/000000.json","/SG_PR_DATA/graphs_sk/05/000006.json"] in eval_pair.py.
The rusult score more than 1.
------------------------------------------------------------------------------console output:
Enumerating unique labels.
I,m confused.
Thank you very much for your open source work,but I do know how to use multiple to train this net,Orz,Please help me T_T
sg_net.py using "cuda:0"
dgcnn.py using "cuda"
your could exchange it with "cuda:"+str(args.gpu_id)
Congratulations, author, great work. I really want to know how to visualize Fig. 7 in this paper? Do you have sample code?
Thanks for your great contribution!
I have a question for testing and training part
You described "We use 1-fold cross-validation and
each sequence is considered as a fold, that is, consider one
sequence as a test set and the others as training sets"
And I understood the LoopClosureDetection based 00,02,05,07,08 sequnce
Could you please tell me in details about which sequnce you used for the 1fold cross validation for testing and training?
Hoping for your reply.
Thanks a lot.
Thanks for the great contribution of this research.
About the visualization part for the dir(data_process),Could you please tell me how could i visualize the .json format 3d data?
by the way,the processed data was classfied into 10 classes for RANGENET++,I would like to know the standard of the classification.
Thanks a lot.
The code is missing
Hello, I have used your semantic mapping solution, starting with pairing to create data. However, the F1 score on my first training attempt was 0.9. May I know what could be the reason? I have replaced the original clustering method.
Hi, I have a question about the pre-processing. In the paper, you process the raw point cloud by segmentation and ignore classes like person, because they are either irrelevant or few in number. So why not ignore the vehicle class, which I think is movable and would interfere with place recognition.
Hello, this work is quite amazing and most of the documentation is very clear. But I am a little confused about how to generate the pair_list in the train folder in the preprocessed data you have provided.
Hello, I would like to ask if the data set of the comparative experiment (such as pointnetvlad) in the paper is also divided according to your script?The pair_list I generated according to this script and I got f1_ score seems that the lowest score is about 0.65, It should not be lower than 0.129 in your paper.
Hello, I'd like to ask how i can get the JSON file format you published in advance. For example, there is pose data in the JSON data you published in advance. You did use pose later, but the data preprocessing code does not involve how to calculate pose ? Can you help me answer my question.Thank you.
Hello,I have benefited a lot from reading your code. How to use the data in rawdata, such as 00_M2DP_db.npy and 00_gt_db.npy in the M2DP00 folder
Hi, thanks for your interesting work! I'm going deep into the code and have a question in line 326 of sg_net.py. The code is as below. In my opinion, batch_feature_1 should be the concatenated feature of one frame. So as batch_feature_2. For what reason makes you append data["features_2"] into batch_feature_1? Tell me if I miss any point. Thanks again for your reply.
for graph_pair in batch:
data = process_pair(graph_pair)
data = self.transfer_to_torch(data, training)
batch_feature_1.append(data["features_1"])
batch_feature_2.append(data["features_2"])
batch_feature_1.append(data["features_2"])
batch_feature_2.append(data["features_1"])
target = data["target"]
batch_target.append(target)
batch_target.append(target)
how to generate text file(like 00.txt) of train?
Hi, I am retraining your model with default settings on a GTX1080 GPU, except the batch size is set to 256, and the nodes used are from semantic Kitti presented by you. When I first test sequence 08, which means sequences 00-07,09,10 as the training sets, the F1 max should achieve 0.900 if I wasn't missing any point. However, the best training result is only around 0.66(F1 max). Could you please give some suggestions on this situation?
I have read your paper in IROS 2020. It is an amazing work!
Could you share the code as soon as possible? I can't wait to reproduce your job now.
Thanks a lot!
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.