Annotation and configuration for few-shot object detection

This is a fork of makesense.ai, with modifications to facilitate few-shot object detection with own datasets. The original make-sense readme with installation instructions can be found here.

This version supports setting up a few-shot training task using our version of the Frustratingly Simple Few-Shot Object framework by Wang et al. This framework needs to be installed to run the training. The generated scripts use a variable FSDET_HOME to refer to the root directory of the framework installation.

• When starting the application, a dialog with few-shot training properties pops up. This dialog can later also be accessed via the application menu. The following properties can be set:
  • the name of the new dataset (the few-shot training problem will then use <basemodel>_<new dataset> for the names of configuration files
  • the base model (COCO60 is the one that can be obtained from the the FsDet model zoo)
  • the name of the annotation file for the new dataset (relative to FSDET_HOME)
  • the image directory (relative to the datasets/ directory of the framework)
• The labels defined will be those used as novel classes.
• Use polygon annotations.
• The exporter will provide three files: the annotations in COCO format, the training configuration in YAML format and the script for running the training. The export dialog provides information where to put these files.
• Optionally, the exporter will write files to a server, specifying the file path relative to the FSDET_HOME root. A REST endpoint needs to be provided (in FewshotExport.ts), to which the file content is to be posted.
• To run the training, make the downloaded script exectuable using chmod +x <filename>.

Acknowledgement

This work has received funding from the European Union’s Horizon 2020 Research and Innovation Programme under grant agreement No 951911 (AI4Media) and from the program ICT of the Future of the Austrian Federal Ministry of Climate Action, Environment, Energy, Mobility, Innovation and Technology (BMK).