akshaytrivedi / driveme

driveME

Tried writing my own MaskRCNN first(implementation in fair-maskrcnn). This is giving errors which are hard to debug and this implementation is very slow.

So moved to matterport Mask RCNN implementation

Using matterport for the competition

1. Run matterport model on simple images (DONE)

2. Run matterport model on Kaggle dataset (DONE)

[Next Steps from here]3. Evaluate the matterport model

Run on test set (Done)

Submit test set predictions to Kaggle (Getting Errors)

Spilt training data into train and validation set (In progress)

Evaluate on validation set and visualize matterport model errors

4. How we can better same matterport model to improve prediction

   Training/Fintuning the model

   Landscapes dataset

   Test image augmentation (May be decreasing size)

   Ensemble of each category i.e we train one classifer for each category vs background

   Ensemble of each video i.e predicting for each image of a video

   Semantic Segmentation for Videos

5. Improving the model architecture