The following segment represents the instance-segmentaion part of the project. You can reproduce our results by following the steps presented below.
The code is based on the following modules and frameworks:
- Python 3.9
- Core: PyTorch, OpenCV, TorchVison
- Functional: NumPy, random, PIL, datumaro, etc.
You can install all the dependencies using the following command:
pip install -r requirements.txtAlso you will need to acquire our dataset to reproduce model training. You can make it with the following command or by follow the disk link.
wget "link"The module's aid is to configure and train the model, as well as saving the metrics, sample images during the way.
- noise_dir - source path to Noise(background) images
- pollen_dir - path to a directory where pollen samples are held
- PATH_TO_MODEL - path, where trained model weights will be saved
- image_save_dir - path, where processed generated images will be saving during the training phase
- log_dir - path to a directory where training metrics will be saved
This module is creating the test dataset and processes it's evaluation. The given results is the IoU metrics: bbox (bounding box), segm (masks)
- test_root - path to CVAT dataset, wich helds the test images and mask polygons for model evaluation
- model_root - path to model weights
This module is inference mode, performing the model's task: instance segmentation (extracting the objects for the passed images).
- path_to_model - path to model weights
- input_dir - path to directory, where target images are held (microscopic images, where the object segmetation is needed)
- output_dir - path to a directory, where finded objects are held, the fromat looks like: sample (1).jpg -> sample(1)_0.jpg, sample(1)_1.jpg...
The following values defines the confidence level, when the model should refer the ecnountered object as a target object (which should be returned)
- score_threshold - (float) overall model's score threshold
- mask_threshold - (float) mask confidence threshold