Comma.ai challenge to predict speed from a dashcam video
Split up the data into train(90%) and validation(10%).
MSE -
- Train - 4.4575
- Validation - 1.089
| Train_graph | Validation Graph |
|---|---|
 |
 |
Key points are tracked using a mask as shown in the sample gif. Rotation of camera is not considered.
Referred to this blog post for some guidance and this implementation for hyperparameters. You can also refer to these slides for the math behind deriving the optical flow parameters.
- - Compute MSE for both train and val
- - Graph for both train and validation
- - Video showing key points