Skip to content

Latest commit

 

History

History
32 lines (23 loc) · 3.15 KB

EXTEND_FUNCTIONALITY.md

File metadata and controls

32 lines (23 loc) · 3.15 KB
Raw

EXTEND FUNCTIONALITY

For most users, if you just want to use the entire set of annotations and skip any additional filtering then skip this section.

The parameter dataset_type can be extended further by modifying the source code. As of now it supports only the kitti type which involves filtering data into easy, moderate and hard categories as defined here.

To add additional parameters in the annotations and different filtering levels for your own custom dataset_type modify the source code in the following places:

The places where the source code should be modified are marked with the comments:

    #EXTEND FUNCTIONALITY HERE
    #UPDATE ANNOTATION MODES HERE

Followed by a short description of the context it needs to updated in.

  1. function name: parse_rec() in $PROJECT_ROOT/dl_algos/py_faster_rcnn/lib/datasets/general_dataset_eval.py This function allows you to add additional annotation data fields to store for a particular image. Add an else if block and check for your custom dataset_type here and add on the additional annotation entries to the structure variable.

  2. function name: determine_sample_mode() in $PROJECT_ROOT/dl_algos/py_faster_rcnn/lib/datasets/general_dataset_eval.py This function allows you to filter data for training and testing and refers to the train_mode and test_mode in the master objdet_experiments_conf.ini file. To extend this, add an else if block and check for your custom dataset_type here and assign a sample_mode (the level) depending on the annotation fields (refer to the kitti part of the code for an example). It is generally more meaningful to make sure that lower levels are subsets of the higher levels. Assign a lower number (sample_mode field) to the lower levels and a higher number to higher levels (Note that this is not compulsory but recommended to make it easier to understand).

  3. function name: general_dataset_eval() in $PROJECT_ROOT/dl_algos/py_faster_rcnn/lib/datasets/general_dataset_eval.py If evaluation needs to be carried out on additional dataset_type, modify this function. There are 2 parts here:

  • ValidInds : Filter out higher level sample modes and get a boolean array for evaluation purposes
  • MinHeight : You can set the minimum height of bounding boxes here (for evaluation)

  1. function name: load_general_dataset_annotation() in $PROJECT_ROOT/dl_algos/py_faster_rcnn/lib/datasets/general_dataset.py For adding additional annotation data for a different dataset_type and for filtering data for training networks with faster-rcnn modify the source code here, keeping in mind the conditions you set earlier on the sample_mode (in parse_rec() and determine_sample_mode()). Keep in mind that you will have to add all annotation fields again even if you have added them in parse_rec().

  2. function name: convert_annotation() in $PROJECT_ROOT/data_utils/pascal_to_yolo.py For filtering data for training networks in darknet (yolo/yolov2) modify the source code here, keeping in mind the conditions you set earlier on the sample_mode (in parse_rec() and determine_sample_mode()).