"BigBalls" Detection And Segmentation Task

Validation Set Run Gif Example

• Here Colored by green just "BigBalls" and by red "TheBiggestBigBalls"

Dataset

In our course team we made up "handmade" markup for dataset by cvat.org We prepare 85 frames for train and augmentation and 10 frames for validation

---

Model

I use RCNN implementation from https://github.com/matterport/Mask_RCNN

"How To"

Train model

• Use CLI command to train from root repo folder

foo@bar:~$ python3 app/model.py --dataset=/path/to/train/dataset

Evaluate the model

• Use CLI command to evaluate the model from root repo folder

foo@bar:~$ python3 app/model.py evaluate --dataset=/path/to/evaluation/dataset

Deploy the model

1. Download repository
2. Download model file from GDrive and unzip to model/ folder
3. Install requirements.txt by pip or conda on you taste

---

Custom metrics

• MSE of the number of abnormal clods.
  • Achieved on val_set MSE = 11.8
• MSE of the size of the biggest one

---

Utilized resources

https://cvat.org/ https://engineering.matterport.com/splash-of-color-instance-segmentation-with-mask-r-cnn-and-tensorflow-7c761e238b46 https://colab.research.google.com/github/RomRoc/maskrcnn_train_tensorflow_colab/blob/master/maskrcnn_custom_tf_colab.ipynb#scrollTo=X7iSzccTL9hM https://github.com/matterport/Mask_RCNN

Creds

• Eugene Borisov, 971901 HITS, email: evgenyboriso@gmail.com