OpenI Dolphin Documentation

Welcome to project Dolphin, which is an open source framework of algorithms using deep learning in several fields of computer vision based on PyTorch: Object Detection, Generative Adversarial Network, Video Action Analysis, Mono Depth Estimation, Activate Learning, Object Tracking and Segmentation, aiming to promote learning of deep-learning based computer vision algorithms and simultaneously simplifies algorithms research experiments.

Note

This documentation contains only API introduction of the project, the installation guidance please check our repo.

Wide Coverage

A variety of computer vision algorithms are integrated, for every included field of them, there exists at least one specific algorithm:

• Object Detection: Faster RCNN

• Video Action Analysis: SlowFast for action recognition, BMN for temporal action localization, MOC-Detector for temporal and spatial action detection

• Generative Adversarial Network: CycleGAN

• Mono Depth Estimation: FCRN

• Activate Learning: Query-based Entropy Sampling

• Object Tracking: FairMOT

• Segmentation: FCN

Modular Design

The workflow of algorithm is separated into several modules: dataset establishing, data augmentation, model building and so on, that is convenient for customization and combination. What’s more, all of setting of modules and hypeparameters can be easily done in a simply configuration file.

Flexible Engine

For adapting to some special algorithms, such as GAN, Activate Learning, a flexible workflow engine is created. It is compatible with controlling sequence of updating different models within a iteration (GAN) and special workflow phase (query-based activate learning algorithms).

The details about this system are provided as below, You can find out more in this documentation.

• Module Configuration

• Module of Model

• Module of Dataset

• Module of Engine