BAR - A Reinforcement Learning Agent for Bounding-Box Automated Refinement

Morgane Ayle, Jimmy Tekli, Julia El Zini, Boulos El Asmar, Mariette AwadAAAI 2020 · DOI

In computer vision, the performance of deep learning models is heavily dependent on the quality and quantity of labeled data. However, generating precise bounding-box (b-box) annotations is a labor-intensive process, often serving as a significant bottleneck in industrial AI pipelines. While much research focuses on improving initial object detection, there is a distinct need for tools that can efficiently correct "noisy" or inaccurate labels.

Ayle et al. (AAAI-20) introduced BAR (Bounding-box Automated Refinement), a reinforcement learning agent designed to automate the fine-tuning of existing bounding boxes. By treating refinement as a sequential decision-making task, BAR can reduce human intervention time by up to 82%.

The Refinement Pipeline

Unlike standard detectors that scan an entire image, BAR starts with a "weak" initial bounding box—perhaps generated by a low-accuracy model or a biased human annotator. The agent then iteratively adjusts the box's boundaries until it tightly encloses the target object.

The BAR agent iteratively refines an inaccurate initial bounding box through a sequence of discrete actions.

Methodology

The authors frame the refinement process as a Sequential Decision Making Problem (SDMP). The agent observes the current state (the image region within and around the box) and selects an action to maximize a cumulative reward.

Action Space and Reward Logic

The agent utilizes a discrete action space consisting of 9 possible moves:

  • Translation: Up, Down, Left, Right.
  • Scaling: Wider, Taller, Fatter, Thinner.
  • Terminal: A "Stop" action triggered when the box is deemed accurate.

The reward function is designed to encourage both accuracy and efficiency. The agent receives a positive reward (+1) for actions that increase the Intersection-over-Union (IoU) with the ground truth and a penalty (-3) for those that decrease it. A significant bonus is awarded for reaching a high IoU threshold ($\beta$) in the fewest possible steps.

Architectural Variants

Two versions of BAR were developed to balance robustness and adaptability:

  1. BAR-DRL (Deep Reinforcement Learning): Utilizes a ResNet50 backbone for high-level feature extraction and a Deep Q-Network (DQN). This version is ideal for complex objects and offline training.
  2. BAR-CB (Contextual Bandits): An online approach using simpler HOG (Histogram of Oriented Gradients) features. It allows for real-time learning and adaptation to new data streams, making it highly effective for consistent industrial parts like stickers.

Experimental Results

The system was evaluated on both a private industrial dataset (car parts and airbag stickers) and the public PASCAL VOC benchmark.

  • Accuracy Improvements: BAR consistently increased the IoU of inaccurate boxes by an average of 0.28. On PASCAL VOC ('aeroplane' class), BAR-DRL improved the Average Precision (AP@50) from 7.9% to 23.5%.
  • Operational Efficiency: In industrial settings, BAR-DRL saved between 30% and 82% of the time typically required for manual correction.
  • Speed: With an inference time of just 15ms per image, the agent adds negligible overhead to labeling workflows.

Conclusion

BAR demonstrates that reinforcement learning is a powerful tool for automating the "polishing" phase of data labeling. By focusing on refinement rather than detection from scratch, the agent operates within a reduced search space, leading to faster convergence and significant reductions in manual labor. Future work aims to integrate BAR into active learning frameworks, where the agent can proactively identify boxes that require human oversight.

Summary updated

BibTeX
@inproceedings{ayle2020bar,
  author = {Morgane Ayle and Jimmy Tekli and Julia El Zini and Boulos El Asmar and Mariette Awad},
  title = {{BAR - A Reinforcement Learning Agent for Bounding-Box Automated Refinement}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i03.5639},
}
Figure