Steps for training with a new dataset

[tgels]

I am trying to train rccn with my own dataset and am trying to document the steps needed:

1. Identify the classes of detection
2. Use the VOC 07 folders as a template or use as it is to start
3. Label the images in VOC 07 format ie. 000001.jpg, etc. and store them in JPEGImages folder
4. Create bounding box annotation for the objects to be detected with the correct labeled image, etc. and store them in the Annotations folder
5. Create training, validation, test data files, class_train.txt, class_test.txt, class_val.txt, class_trainval.txt (see ImageSets/Main folder for examples) -- Use the same format [ imageid -1 (false) or imageid 1 (true) or imageid 0 (not sure but seems to be true)]

Do we need the others ? like ImageSet/Layout, ImageSet/Segmentation ?

Also, do we need the SegmentationClass and SegmentationObjects folders ? Any ideas on how these were created ? Here is an example:

[wuxinhong]

how do you Create training, validation, test data files, class_train.txt, class_test.txt, class_val.txt, class_trainval.txt??

[tgels]

If you have a list of images and their categories, you use it to create the class_train.txt, class_test.txt, class_val.txt, and class-trainval.txt. Divide the imageset into training images, test images, and validations images.

The different ids go into class_train.txt, etc. The format of the class_train.txt is in the form "key" "value". Key is the image id, and value is either 0, 1, -1. -1 if it does not belong to the class and 1 if it belongs to that class.

For eg: bike_train.txt 00001 -1 00002 1

where 00002 is 00002.jpg and belongs to the bike class.

[sreevasu]

HI,

Please can you help me on how to create bounding box annotation for my own dataset.