Code for AAAI2024 paper: Towards Evidential and Class Separable Open Set Object Detection
The Evidential Object Detector (EOD) is implemented using the Detectron2 object detection framework and follows the open-set configuration of Opendet. Gratitude is extended to the original authors for their valuable contributions and commitment to open source.
The fundamental setup tasks (e.g., installation and dataset preparation) can be easily accomplished by referring to the guidelines provided in the two aforementioned projects.
The essential code from the paper is provided. This enables straightforward reproduction through the following steps.
-
Identify the modules that need to be replaced in the Faster-RCNN according to the detection framework and open-set settings.
-
Copy the corresponding files or folders from this repository to the appropriate locations in the project.
-
Make modifications and adaptations as needed.
The training process is the same as Detectron2 and Opendet.
python tools/train_net.py --num-gpus 8 --config-file configs/faster_rcnn_R_50_FPN_3x_EOD.yamlRun the following command for testing.
python tools/train_net.py --num-gpus 8 --config-file configs/faster_rcnn_R_50_FPN_3x_EOD.yaml --eval-only MODEL.WEIGHTS output/faster_rcnn_R_50_FPN_3x_EOD/model_test.pth