This repo has code for running continual object detection experiments on top of the Detectron2 library. It is fully written outside Detectron, which allows easy modifiation and upgrades to newer, future versions of Detectron. This repo serves three main purposes: continual object dataloaders, an implemenation and improvement of FasterILOD, and extensive logging capabilities for analysis of object detectors.
The data-sets module contains dataloaders and integration with Detectron's dataloaders. See here for a list of all registered datasets. For VOC, it's easy to define new splits on the VOC2007 benchmark. COCO is currently limited to the 40+40 setting, but this will be upgraded soon. The SODA10M are those defined for the ICCV '21 challenge, see [TODO].
The ilod.py file is both an example of how to implement COD methods without changing the Detectron2 library,
a reimplemenation of FasterILOD, and an improvement thereof. See the
documentation of the file for more details. train.py contains logic to train and test such a model. The
config folder contains example configs for VOC10+10. The scripts folder contains an example
script for the training thereof.
log_utils.py implements a decorator for the inference call on a GeneralizedRCNN model of detectron. It has two main
capabilities: (1) it creates a hook of the RPN, which logs all proposals to a file on inference. This allows for a
separate analysis of the RPN within a two-stage detector, since its performance is hard to judge given only the
predictions of the ROI-heads. (2) Seperate evaluation of the ROI-heads, by evaluating the ground truth boxes of each
image and log those to a file. Both files can then be analyzed with the analysis module. (Note:
the code of the analysis module is not well documented yet, but that will come soon). These logging functionalities
are separate from continual detection, and can thus be used with any GeneralizedRCNN model.