dli2016 / isee-algorithm-interfaces-py Goto Github PK

This project defines the interfaces of the vision algorithms integrated in ISEE platform (python version). The Intelligent Scene Exploration and Evaluation (ISEE) platform is a general computer vision system for large-scale video parsing. The users can upload their custom vision algorithms and design the execution plan flexibly. To achieve this, we release the interfaces of different vision algorithms, so that the user can run their algorithms on ISEE conveniently.

• common: the file isee_interface.py in this folder define the abstract class ISEEVisAlgIntf that all the other vision algorithm classes should inherent.
• demo: it contains all the main files to run the example algorithms that achieve by inherenting the abstract class ISEEVisAlgIntf.
• densepose_estimation: the module for densepose estimation.
  • methodname_version: the root folder for the user uploaded algorithms. the user should name the folder through combining the method name and version flag.
    • DenseposeEstimation.py: complete the ISEEDenseposeEstimation by implement the methods in ISEEVisAlgIntf.
    • dependencies: all the dependent codes or packages to achieve densepose estimation.
  • densepose_rcnn_coco: it provides an example to densepose estimation using the projects of Detection2 and DensePose.
• instance_segmentation: the module for instance segmentation (its orgnization is similar with densepose_estimation).
• keypoints_detection: the module for keypoints detection including person pose and face.
  • pose: person pose etimation (its orgnization is similar with densepose_estimation).
  • face: face keypoints detection (comming soon ...).
• object_detection: the module for object detection (its orgnization is similar with densepose_estimation).
• panoptic_segmentation: the module for panoptic segmentation (its orgnization is similar with densepose_estimation).
• reid: the module for extracting reid features (comming soon ...).
• attribute_recognition: the module for attribute recognition (comming soon ...).
• tracking: the module for object tracking (comming soon...).
• action_recognition: the module for action recognition (comming soon ...).

• To run the custom algorithms on ISEE, only a cofiguration file (e.g., XML, JSON, YAML, etc.) can be provided to initialize the method.
• The output formats of different algorthims have speficied in each module (more details can be found in the README of each module). Please do not change the format which will influence the data management in ISEE.
• More vision algorithms will be introduced in the future.

• Detectron2
• DensePose

@article{li2019isee,
  title={ISEE: An Intelligent Scene Exploration and Evaluation Platform for Large-Scale Visual Surveillance},
  author={Li, Da and Zhang, Zhang and Yu, Kai and Huang, Kaiqi and Tan, Tieniu},
  journal={IEEE Transactions on Parallel and Distributed Systems},
  volume={30},
  number={12},
  pages={2743--2758},
  year={2019},
  publisher={IEEE}
}