iMaterialist 2020 - 1st Place Solution

Model architecture:

• Mask R-CNN model
• SpineNet-143 + FPN backbone
• An extra head to classify attributes

Training:

• Pre-trained on the COCO dataset
• Image resolution: 1280x1280
• Focal loss for the attributes head
• Augmentations: random scaling (0.5x - 2.0x), v3 policy from the AutoAugment (modified to support masks)

All the changes were made on top of the TPU Object Detection and Segmentation Framework.

Read more about the solution in the Kaggle post.

Download the model weights here.

Cloud TPUs

This repository is a collection of reference models and tools used with Cloud TPUs.

The fastest way to get started training a model on a Cloud TPU is by following the tutorial. Click the button below to launch the tutorial using Google Cloud Shell.

Note: This repository is a public mirror, pull requests will not be accepted. Please file an issue if you have a feature or bug request.

Running Models

To run models in the models subdirectory, you may need to add the top-level /models folder to the Python path with the command:

export PYTHONPATH="$PYTHONPATH:/path/to/models"