Linux Install faster_rcnn in TNT (Matlab wrapper by ShaoqingRen):

Step1: go into your path

cd /home/Username

Step2: download faster_rcnn from github

$ git clone --recursive https://ShaoqingRen/faster_rcnn.git

Step3: Compile file Makefile.config

$ cd external/caffe

If there is no files of caffe, just download Caffe as before (Linux Install Caffe in TNT).

$ cp Makefile.config.example Makefile.config $ vim Makefile.config modify: MATLAB_DIR := as MATLAB_DIR := /usr/local/MATLAB/R2014a

Tipp: It's better use Matlab_R2014a, I have tryed R2016b, there is many bugs.

Step4: modify V1LayerParameter Layer type:

find the file upgrade_proto.cpp: $ cd /home/Username/faster_rcnn/external/caffe/src/caffe/util modify upgrade_proto.cpp: $ vim upgrade_proto.cpp after case V1LayerParameter_LayerType_THRESHLOD: return "Threshold";

add:

case V1LayerParameter_LayerType_RESHAPE:
	return "Reshape";
case V1LayerParameter_LayerType_ROIPOOLING:
	return "ROIPooling";
case V1LayerParameter_LayerType_SMOOTH_L1_LOSS:
	return "SmoothL1Loss"; 

then Esc add : wq return

Step5. Install Caffe to local

$ cd caffe $ make clean $ make all -j16 $ make test -j16 $ make runtest -j16 $ make matcaffe Install matlab API, become MEX File $ make pycaffe Install python API

Step6: Download Pre-trained Model:

in MATLAB run: $ run fetch_data/fetch_faster_rcnn_final_model.m

Step7: run faster_rcnn in Matlab:

$ run faster_rcnn_build.m and $ run startup.m

Step8: test faster_rcnn:

$ run experiments/script_faster_rcnn_demo.m

Step9: Preparation for Training:

1. Run fetch_data/fetch_model_ZF.m to download an ImageNet-pre-trained ZF net.
2. Run fetch_data/fetch_model_VGG16.m to download an ImageNet-pre-trained VGG-16 net.
3. Download VOC 2007 and 2012 data to ./datasets

Step10: Training a model

1. Run experiments/script_faster_rcnn_VOC2007_ZF.m to train a model with ZF net. It runs four steps as follows:

• Train RPN with conv layers tuned; compute RPN results on the train/test sets.
• Train Fast R-CNN with conv layers tuned using step-1 RPN proposals; evaluate detection mAP.
• Train RPN with conv layers fixed; compute RPN results on the train/test sets.
• Train Fast R-CNN with conv layers fixed using step-3 RPN proposals; evaluate detection mAP.
• Note: the entire training time is ~12 hours on K40.

2. Run experiments/script_faster_rcnn_VOC2007_VGG16.m to train a model with VGG net.

• Note: the entire training time is ~2 days on K40.

3. Check other scripts in ./experiments for more settings.

Tipps:

1. The first time I run the script_faster_rcnn_VOC2007_ZF.m, ther is a error about text_read, I think there is something wrong with datasets. You can fing in about 28th and 29th rows of script_faster_rcnn_VOC2007_ZF.m , it use voc2007_trainval. Open this file and modify about 11th and14th rows, trainval to train(or the other name, the train model in your model's name /faster_rcnn/datasets/VOCdevkit2007/VOC2007/ImageSets/Main/*.txt/). And then try again, it works.

2. when you run train Model in rcnn (Linux OX), there is a Error: Check failure: fd != -1 (-1 vs. -1) File not find: .\models\rpn_prototxts\ZF\train_val.prototxt Because .\ is not a valid path in Ubuntu (Server athene of TNT use Ubuntu OX ). Solution: change all the \\ to / in ./faster_rcnn/models/rpn_prorotexts/*.prototxt
   Attetion： modify every file under ./faster_rcnn/models/faster_rcnn_prototxts und ./faster_rcnn/models/rpn_prororxts

Step11. Training your own model:

1. Build your own Dataset (like VOCdevkit2007)

My Datasets is NYU_deep data, but in faster_rcnn we have to train the dataset, which like VOC2007 format -- Image name (000001.jpg), Boundingbox information -- .xml files. ther is three Conversion script to build Datasets: datasets_tools:

• creat_traintxt.m (link: http://blog.csdn.net/u014696921/article/details/52950218)
• creat_xml.m (by liao)
• png_jpg.m (by wei)

2. Build files in faster_rcnn

   • In VOCdevkit2007/result build a file with your Datasets' name like NYU_rcnn, and under this file build a file Main
   • In VOCdevkit2007/local build a file with your Datasets' name like NYU_rcnn

3. Modify code

   • In datasets/VOCdevkit2007/VOCcode/VOCinit.m modify VOC2007 to your datasets' name and write your training classes as plane

here you can know the number of your classes K

- In `functions/fast_rcnn/fast_rcnn_train.m`

modify val_iters to 1/5 of your val - funvtions/rpn/proposal_train.m same as before - In imdb/imdb_eval_voc.m modify to:

4. Modify model
   • In models/fast_rcnn_prototxts/ZF/train_val.prototxt K is number of classes in your own datasets
   • In models/fast_rcnn_protoxtxs/ZF/test.prototxt same as before
   • In models/fast_rcnn_prototxts/ZF_fc6/train_val.prototxt
   • In models/fast_rcnn_prototxts/ZF_fc6/test.prototxt

Attention:

1. If there is bug in your training, before start restart, don't forget delete old output and imdb/cache
2. xml must be same format as VOC2007(./datasets/VOCdevkit2007/VOC2007Annotations) spacecan not instead tabel
3. Modify the number of Iterations: in ./experiments/+Model/ZF_for_Faster_RCNN_VOC2007.m the code solver_30k40k that menas use which file of Interations, you can open ./models/fast_rcnn_prototxts and ./models/rpn_prototxts build new .prototxts for your own Iterations.
4. In imdb\imdb_eval_voc.m modify do_eval = (str2num(year) <= 2007) | ~strcmp(test_set,'test'); to %do_eval = (str2num(year) <= 2007) | ~strcmp(test_set,'test'); do_eval = 1;
5. 2 usefull blogs for training faster_rcnn link: http://blog.csdn.net/sinat_30071459/article/details/50546891 http://blog.csdn.net/u014696921/article/details/52950218

5. Run experiments/script_faster_rcnn_VOC2007_ZF.m

stay your path just in 'faster_rcnn', not in 'experiments'

Step12. Test result of your own model

1. After training, at first open output/faster_rcnn_final/faster_rcnn_VOC2007_ZF/detection_test.prototx delete layers until relu5（include relu5) and then modify the data under data and in layer roi_pool5 modify :

2. Test:

• Open \experiments\script_faster_rcnn_demo.m.
• modify to your own model as:
• modify to your own Images:
• If the number of classes of your own datasets more than VOC2007, modify showboxes to: then you can run and test your model.