Uncertainty-aware human-machine collaboration in Camouflaged Object Detection.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Biomedical Engineering

Background:

• Camouflaged Object Detection (COD) is crucial for applications requiring identification of concealed objects.
• Existing methods often struggle with complex camouflage, necessitating improved reliability and efficiency.
• Human-machine collaboration offers a promising avenue to leverage complementary strengths.

Purpose of the Study:

• To develop a novel human-machine collaboration framework for Camouflaged Object Detection (COD).
• To enhance COD performance by integrating computer vision (CV) models with noninvasive brain-computer interfaces (BCIs).
• To improve system reliability and reduce human cognitive load in complex detection tasks.

Main Methods:

• Proposed a multiview backbone for estimating uncertainty in CV predictions.
• Utilized uncertainty estimation during training to boost efficiency.
• Implemented a human evaluation system using RSVP-based BCIs for low-confidence predictions during testing.

Main Results:

• Achieved state-of-the-art results on the CAMO dataset, improving balanced accuracy (BA) by 4.56% and F1 score by 3.66%.
• Demonstrated significant performance gains for top participants (up to 7.6% BA, 6.66% F1).
• Confirmed a strong correlation between model confidence and detection precision.

Conclusions:

• The proposed framework effectively enhances Camouflaged Object Detection through human-machine collaboration.
• The integration of uncertainty-aware CV and BCIs improves system reliability and efficiency.
• This approach provides a foundation for advanced real-world COD applications and human-computer interaction.