TrashTron2

TrashTron2 is a Instance-Segmentation-Model based on FAIR's Detectron2. TrashTron2 is currently trained on over 200+ Images of trash provided by TACO.

You can use and train TrashTron2 in the associated TrashTron2 Notebook with Google Collab. See the instruction under GettingStarted.

Outcome

Models

• ResNext-101
• ResNet-101
• ResNet-50

Scores

Model	AP	AP50	AP75
ResNet-50	58.314	73.825	61.755
ResNet-101	59.584	76.967	62.230
ResNext-101	59.305	78.326	63.566

Accuracy

False Negatives

False Positives

To-Do's

• release v0.1

• Increase number of annotated images

• fine tune models on new images

• release v0.2

• deploy Resnet-50 on NVIDIA JetsonNano