2019/08/01 Jiadai Sun [email protected]

Query Point Cloud	Dataset Model	Result to be modified

If the information of prediction category is used, the result is good(As shown in the figure above), but if it is searched directly in the database, the effect is not perfect and needs to be improved.

• Making ObjectNN HDF5 Data Set

  The part of how to deal with SHREC17-ObjectNN, can find in https://github.com/MaxChanger/DealWithSHREC

• Add query_objectnn_clf.py

  Extract and Used the Global Feature [1024] as the Feature of Query & dataset

  Save the result to GlobalFeature/database0.csv ... database103.csv & query0.csv ... query13.csvq

• Add calcSimilarity.py

  Load database.csv & query.csv , use cosine_similarity to calculate similarity between query and dataset. To be optimized.

• Add copyAndViewResult.py

  According to similarity, we can find the highest similarity model in database, copy them to this folder

  ---Result
     |--query201_Chair_OfficeChair
     	  |--036_529b.ply
     	  |--model
     	  	 |--01_Chair_OfficeChair.obj
     	  	 |--...
     	  	 |--10_Chair_OfficeChair.obj  (Top-K)

• Training model in ModelNet40 and ObjectNN

  Reimplemented ModelNet40 -- the accuracy of classification 89.18%

  I use objecnn20_data_hdf5_2048 made by myself ( SHREC'17 Objec) -- the accuracy of classification 64.16%, if use the pretrained model in ModelNet40, can improve to 66.36%

  And I use to Global Feature generate by pointnet-0.663555-0082.pth to retrieve model

• Training Semantic Segmentation in indoor3d_sem_seg_hdf5_data

  This part of the work was done by the original author. I just reimplemented it again.

• Some *.sh to help me

• If you query directly in the database, the result is not perfect. But if I use category information as a priori information, I can get good results (the first picture)
• Query speed needs to be improved
• The way you view results can be optimized

The following parts are original README.md

This repo is implementation for PointNet and PointNet++ in pytorch.

• Download ModelNet here for classification and ShapeNet here for part segmentation. Uncompress the downloaded data in this directory. ./data/ModelNet and ./data/ShapeNet.
• Run download_data.sh and download prepared S3DIS dataset for sematic segmantation and save it in ./data/indoor3d_sem_seg_hdf5_data/

• python train_clf.py --model_name pointnet

• python train_clf.py --model_name pointnet2

• Training Pointnet with 0.001 learning rate in SGD, 24 batchsize and 141 epochs.
• Training Pointnet++ with 0.001 learning rate in SGD, 12 batchsize and 45 epochs.

• python train_partseg.py --model_name pointnet

• python train_partseg.py --model_name pointnet2

• Training both Pointnet and Pointnet++ with 0.001 learning rate in Adam, 16 batchsize, about 130 epochs and 0.5 learning rate decay every 20/30 epochs.
• Class avg is the mean IoU averaged across all object categories, and inctance avg is the mean IoU across all objects.
• In official version PointNet, author use 2048 point cloud in training and 3000 point cloud with norm in testing. In official version PointNet++, author use 2048 point cloud with its norm (Bx2048x6) in both training and testing.

• python train_semseg.py --model_name pointnet

• python train_semseg.py --model_name pointnet2

• Training Pointnet with 0.001 learning rate in Adam, 24 batchsize and 84 epochs.
• Training Pointnet++ with 0.001 learning rate in Adam, 12 batchsize and 67 epochs.

cd visualizer
bash build.sh #build C++ code for visualization

• PointNet and PointNet++
• Experiment
• Visualization Tool

halimacc/pointnet3
fxia22/pointnet.pytorch

Official PointNet and Official PointNet++