Rotated detections

Version 0.2.0 introduces rotated detections.

Added

• train arguments:
  • --rotated-bbox: Trains a model is predict rotated bounding boxes [x, y, w, h, theta] instead of axis aligned boxes [x, y, w, h].
• infer arguments:
  • --rotated-bbox: Infer a rotated model.

Changed

The project has reverted to the name Object Detection Toolkit (ODTK), to better reflect the multi-network nature of the repo.

• retinanet has been replaced with odtk. All subcommands remain the same.

Limitations

• Models trained using the --rotated-bbox flag cannot be exported to ONNX or a TensorRT Engine.
• PyTorch raises two warnings which can be ignored:

Warning 1: NCCL watchdog

[E ProcessGroupNCCL.cpp:284] NCCL watchdog thread terminated

Warning 2: Save state warning

/opt/conda/lib/python3.6/site-packages/torch/optim/lr_scheduler.py:201: UserWarning: Please also save or load the state of the optimzer when saving or loading the scheduler.
  warnings.warn(SAVE_STATE_WARNING, UserWarning)

Augmentation

This release adds image augmentation (brightness, contrast, hue, saturation) and four degree rotational augmentation.

Added parameters:

• --augment-rotate: Randomly rotates the training images by 0°, 90°, 180° or 270°.
• --augment-brightness (float): Randomly adjusts brightness of image. The value sets the standard deviation of a Gaussian distribution. The degree of augmentation is selected from this distribution. Default: 0.05
• --augment-contrast (float): Randomly adjusts contrast of image. The value sets the standard deviation of a Gaussian distribution. The degree of augmentation is selected from this distribution. Default: 0.05
• --augment-hue (float): Randomly adjusts hue of image. The value sets the standard deviation of a Gaussian distribution. The degree of augmentation is selected from this distribution. Default: 0.01
• --augment-saturation (float): Randomly adjusts saturation of image. The value sets the standard deviation of a Gaussian distribution. The degree of augmentation is selected from this distribution. Default: 0.05
• --regularization-l2 (float): Sets the L2 regularization of the optimizer. Default: 0.0001

retinanet-examples 19.04

This pre-release, corresponding with the NVIDIA GPU Cloud (NGC) PyTorch 19.04 container version, includes the first iteration of pretrained RetinaNet models created with this project:

• ResNet18FPN backbone
• ResNet34FPN backbone
• ResNet50FPN backbone
• ResNet101FPN backbone
• ResNet152FPN backbone