Final for ECE 498 SM Nathan Cueto - Fengmao Zheng

We propse some changes to further optimize the Mask-RCNN COCO performance

• Dataset: Reduce the 2014 dataset to only include relevant, outdoor images. Also, reduce number of classes from 80 to 12 classes. Lesser classes should mean a smaller network.

• Backbone: Smaller network should allow for the lower throughput of a cheaper feature-extraction backbone - resnet50.

• Image augmentation: Implement various image augmentation techniques for training

• Evaluation: Interpolate framerate: Only run on a fraction of input frames to prevent redundant inferences.

Most code is in samples/final/coco.py. We also used the .ipynb notebooks for further verification during the training process.

The training notebook is linked as 498sm_train_final.ipynb.

Results are posted on YouTube here: