-
ResNet50 and Xception were too heavy for this assignment, i found that out by training them for several epochs and my training accuracy varying getting close to 90% and validation still hanging at 20% this happened within 15 epochs for Xception.
-
Tried to extract the images from the given bounding box data and then resized them, instead of getting a higher accuracy i succumbed to worse results and inferred that maybe background here was substantial information. Wasted a lot of days in figuring out why this did not work.
-
Added custom image augmentation for creating random black blobs in images but did not give substantial increase in accuracy results were pretty much the same.
-
Best validation accuracy achieved was 51% and cyclic LR (custom function again) was used.
- Cyclic LR
- Try multiple Architectures.
- Custom Augmentation
- Try Focal Loss
- Make a custom architecture.