TPU Object Detection and Segmentation Framework provides implementations of common image classification, object detection and instance segmentation models in Tensorflow. Our models produce the competitive results, can be trained on multiple platforms including GPU and TPUs, and have been highly optimized for TPU performance. It also features latest research including Auto-Augument, NAS-FPN, ShapeMask, and SpineNet.

** Instance segmentation results of our Mask R-CNN model.

• May 3, 2020: Update inference latency on V100/P100 GPUs for RetinaNet models in MODEL_ZOO.md.
• April 10, 2020: Launch the new README.md, GETTING_STARTED.md, and MODEL_ZOO.md. Release initial models.

• Tasks:
  • Image classification
  • Object detection
  • Instance segmentation
• Meta-architectures:
  • RetinaNet
  • Faster / Mask R-CNN
  • ShapeMask
• Backbones:
  • ResNet
  • SpineNet
• Feature pyramids:
  • FPN
  • NAS-FPN
• Other model features:
  • Auto-Augment
• Training platforms:
  • Single machine GPUs
  • Cloud TPU
  • Cloud TPU Pods

MODEL_ZOO.md provides a large collection of baselines and checkpoints for object detection, instance segmentation, and image classification.

Please follow the instructions in GETTING_STARTED.md.