Home

Awesome

Scripts for SHREC2017 Track3 using RotationNet

Getting Started

1. Prepare caffe-rotationnet2

$ git clone https://github.com/kanezaki/caffe-rotationnet2.git  
$ cd caffe-rotationnet2

Prepare your Makefile.config and compile.

$ make; make pycaffe

2. Download initial weights for fine-tuning the models

Please download the file "ilsvrc_2012_train_iter_310k" according to R-CNN repository
This is done by the following command:

$ wget http://www.cs.berkeley.edu/~rbg/r-cnn-release1-data.tgz  
$ tar zxvf r-cnn-release1-data.tgz

3. Render multi-view images

First, download the 3D model data from SHREC2017 Track3 competition website.
Unzip test_normal.zip, test_perturbed.zip, train_normal.zip, train_perturbed.zip, val_normal.zip, and val_perturbed.zip in data/ directory.
Then, do the following processes in matlab.

$ matlab  
>> render_SHREC('test_normal');  
>> render_SHREC('test_perturbed');  
>> render_SHREC('train_normal');  
>> render_SHREC('train_perturbed');  
>> render_SHREC('val_normal');  
>> render_SHREC('val_perturbed');

4. Train RotationNet models

$ tar zxvf TXT.tar.gz  
$ ./caffe-rotationnet2/build/tools/caffe train -solver prototxt/rotationnet_alex_SHREC2017_case2_solver_b900.prototxt -weights ilsvrc_2012_train_iter_310k 2>&1 | tee log_SHREC2017_alex.txt  
$ ./caffe-rotationnet2/build/tools/caffe train -solver prototxt/rotationnet_alex_SHREC2017_case2_solver_b900_perturbed.prototxt -weights ilsvrc_2012_train_iter_310k 2>&1 | tee log_SHREC2017_alex_perturbed.txt  
$ ./caffe-rotationnet2/build/tools/caffe train -solver prototxt/rotationnet_alex_SHREC2017_SUB_case2_solver_b900_ratio.prototxt -weights ilsvrc_2012_train_iter_310k 2>&1 | tee log_SHREC2017_SUB_alex_ratio.txt  
$ ./caffe-rotationnet2/build/tools/caffe train -solver prototxt/rotationnet_alex_SHREC2017_SUB_case2_solver_b900_ratio_perturbed.prototxt -weights ilsvrc_2012_train_iter_310k 2>&1 | tee log_SHREC2017_SUB_alex_ratio_perturbed.txt

You probably have to set base_lr = 0.0005 for the first few iterations (~1000) and then increase it to base_lr = 0.001.
Or, you can also download pretrained models here.

$ cd models/  
$ wget https://www.dropbox.com/s/civfcz97xh5oy7l/rotationnet_alex_SHREC2017_case2_b900_iter_68000.caffemodel  
$ wget https://www.dropbox.com/s/yqev08sn1yl4xib/rotationnet_alex_SHREC2017_case2_b900_perturbed_iter_75000.caffemodel  
$ wget https://www.dropbox.com/s/7c77rczf231uf7q/rotationnet_alex_SHREC2017_SUB_case2_b900_ratio_iter_78000.caffemodel  
$ wget https://www.dropbox.com/s/lgtvfgj9l7ck5mu/rotationnet_alex_SHREC2017_SUB_case2_b900_ratio_perturbed_iter_78000.caffemodel

5. Save scores

$ bash save_scores_alex.sh train normal  
$ bash save_scores_alex.sh train perturbed  
$ bash save_scores_alex.sh val normal  
$ bash save_scores_alex.sh val perturbed  
$ bash save_scores_alex.sh test normal  
$ bash save_scores_alex.sh test perturbed  
$ bash save_scores_alex_ratio_subclass.sh train normal  
$ bash save_scores_alex_ratio_subclass.sh train perturbed  
$ bash save_scores_alex_ratio_subclass.sh val normal  
$ bash save_scores_alex_ratio_subclass.sh val perturbed  
$ bash save_scores_alex_ratio_subclass.sh test normal  
$ bash save_scores_alex_ratio_subclass.sh test perturbed

6. Do retrievals

Download train.csv and val.csv.

$ wget http://vision.princeton.edu/ms/shrec17-data/train.csv  
$ wget http://vision.princeton.edu/ms/shrec17-data/val.csv

Main results

$ python classify_dir_and_save.py classes.txt results_alex train normal Alex  
$ python classify_dir_and_save.py classes.txt results_alex train perturbed Alex  
$ python classify_dir_and_save.py classes.txt results_alex val normal Alex  
$ python classify_dir_and_save.py classes.txt results_alex val perturbed Alex  
$ python classify_dir_and_save_test.py classes.txt results_alex normal Alex  
$ python classify_dir_and_save_test.py classes.txt results_alex perturbed Alex

Subclass results

$ python classify_dir_and_save.py classes.txt results_alex_ratio_subclass train normal AlexRatio_sub 203  
$ python classify_dir_and_save.py classes.txt results_alex_ratio_subclass train perturbed AlexRatio_sub 203  
$ python classify_dir_and_save.py classes.txt results_alex_ratio_subclass val normal AlexRatio_sub 203  
$ python classify_dir_and_save.py classes.txt results_alex_ratio_subclass val perturbed AlexRatio_sub 203  
$ python classify_dir_and_save_test.py classes.txt results_alex_ratio_subclass normal AlexRatio_sub 203  
$ python classify_dir_and_save_test.py classes.txt results_alex_ratio_subclass perturbed AlexRatio_sub 203

Results with subclass reranking

$ python rerank_dir_and_save.py Alex AlexRatio_sub AlexSR

We submitted AlexSR for SHREC2017 competition. However, Alex is actually better.