PeekingDuck Versions Save

Documentation

• Add typeguard to macOS installation instructions (https://github.com/aisingapore/PeekingDuck/pull/695)

Process

• Change --verify_install from a runtime option to a CLI command verify-install.
• Add --viewer to activate the new PeekingDuck Viewer (see User Interface below).

Features

• Add new augment.undistort node to remove distortion from a wide-angle camera image.
• Add new dabble.camera_calibration node to calculate the camera coefficients for removing image distortion, used by augment.undistort.
• Add new instance segmentation model, with new model nodes model.mask_rcnn and model.yolact_edge. model.mask_rcnn supports ResNet50 and ResNet101 backbones. model.yolact_edge supports ResNet50, ResNet101 and MobileNetV2 backbones.
• Add new draw.mask node to draw instance segmentation masks on image, can be used to mask out objects or background.
• Add new TensorRT optimized models for model.movenet and model.yolox.
• Add node config type checking for all pipeline nodes with user configurable parameters. Will throw a runtime error on wrong config type.

Deprecations and Removals

• peekingduck run --verify_install is deprecated and replaced by peekingduck verify-install command instead.

Documentation

• Add new use case Privacy Protection (People & Screens) using instance segmentation and blurring.
• Add new Edge AI documentation on how to install and run TensorRT models, including performance benchmark charts.

Dependencies

• Add typeguard library typeguard ≥ 2.13.3.

User Interface

• Add PeekingDuck Viewer, a GUI built on TkInter that supports a playlist of multiple pipelines, callable via peekingduck run --viewer. Upon completion of a pipeline, the user may replay the output video or scrub to a specific frame of interest for analysis.

Refactor

• Streamline peekingduck/cli.py by encapsulating source codes for CLI commands nodes, init, run and create-node under peekingduck/commands/ folder.
• Add PeekingDuckLogo.png to setup.cfg and add setup.py to support older pip versions.
• Add new model config field model_format for model.movenet and model.yolox to allow selection between the original models or the new TensorRT models.
• Refactor model.movenet and model.yolox inference code to work with TensorRT models.
• Define _get_config_types() method for all nodes with user configurable parameters. Relevant to most nodes under peekingduck/pipeline/nodes/ folder.
• Use ThresholdCheckerMixin to check bounds in dabble.tracking.

Third release candidate for 1.3.0 (third time's a charm as they say)

Second release candidate for 1.3.0.

First release candidate for 1.3.0.

Bug fixes

• Fix RecursionError when running peekingduck --verify_install on Windows (https://github.com/aisingapore/PeekingDuck/pull/688)

First release candidate for 1.2.3

Process

• Rename aimakerspace to aisingapore (https://github.com/aisingapore/PeekingDuck/pull/684)

Features

• Precompute draw.legend legend box width to fit content and add box_opacity config option (https://github.com/aisingapore/PeekingDuck/pull/660, https://github.com/aisingapore/PeekingDuck/pull/677)
• --config_path now supports absolute paths (https://github.com/aisingapore/PeekingDuck/pull/665)

Deprecations and Removals

• Model node config option detect_ids has been changed to detect (https://github.com/aisingapore/PeekingDuck/pull/653)
• Fix MOSSE tracker in dabble.tracking to only be initialized when there is a detection (https://github.com/aisingapore/PeekingDuck/pull/662)
• Ensure saved_video_fps in the input.visual node is within the range of values supported by cv2.VideoWriter (https://github.com/aisingapore/PeekingDuck/pull/664)
• Fix output.screen window size to adopt input video frame size when do_resizing=False (https://github.com/aisingapore/PeekingDuck/pull/669)

Dependencies

• Tighten opencv-contrib-python version to satisfy linting requirements (https://github.com/aisingapore/PeekingDuck/pull/657)

First release candidate for 1.2.2

Process

• Models now download individual weights files for each model type instead of an entire folder of all model types
• Limit custom node directories to be created within src/

Bug Fixes

• Fix input.visual issue where filename data type is not set if the source is a single image file
• Fix input.visual issue where approximate progress shows 10% even when the folder has been fully processed

Improved Documentation

• Add tutorial for using own models

Refactor

• Redirect tqdm (used when downloading weights) to use stdout instead of stderr
• Implement mixins for threshold checking and weights downloading