Awesome

Interactron: Embodied Adaptive Object Detection

By Klemen Kotar and Roozbeh Mottagh

Interactron is a model for interactive, embodied object detection. It is the official codebase for the paper Interactron: Embodied Adaptive Object Detection. Traditionally object detectors are trained on a fixed training set and frozen at evaluation. This project explores methods of dynamically adpating object detection models to their test time environments using MAML style meta learning and interactive exploration.

Setup

• Clone the repository with git clone https://github.com/allenai/interactron.git && cd interactron.

• Install the necessary packages. If you are using pip then simply run pip install -r requirements.txt.

• If running on GPUs, we strongly recommend installing PyTorch with conda.

• Download the pretrained weights and data to the interactron directory. Untar with

tar -xzf pretrained_weights.tar.gz
tar -xzf data.tar.gz

Results

Bellow is a summary of the results of the various models.

For more detaile results please see the full paper Interactron: Embodied Adaptive Object Detection.

Evaluation

Evaluation of the Interactron model can be performed by running python evaluate.py --config=configs/interactron.yaml. The code will automatically take over any available GPUs. Running the evaluation on a CPU could take several minutes. The evaluator will output visualizations and results in a folder called evaluation_results/. To evaluate other models, select one of the other config files in configs/.

Training

Training of the Interactron model can be performed by running python train.py --config=configs/interactron.yaml. The code will automatically take over any available GPUs. To train using the default configuration, at least 12GB of VRAM is necessary. Training takes roughly five days on a high performance machine using a RTX 3090 GPU. The trainer will output results in a folder called training_results/. To train other models, select one of the other config files in configs/.

Citation

@inproceedings{kotar2022interactron,
  title={Interactron: Embodied Adaptive Object Detection},
  author={Klemen Kotar and Roozbeh Mottaghi},
  booktitle={CVPR},  
  year={2022},
}

Parts of the codebase were derived from other repositories and modified (like the DETR model code) and have a crediting comment on the first line of the file.