Bias-free object detection
Current object detection methods suffer from the problem of dataset bias, as apparent in Computer Vision in general. For companies deploying these methods around the world to commercial products this is a significant problem. Recent work explores different approaches to reducing bias in the resulting networks, e.g. by penalizing overfitting on biased correlations. This however does not extend to object detection, where the network architectures are more involved and biased representations might not be apparent from the network outputs.
This research would explore methods to reduce unwanted dataset bias in object detection.
- How does one optimize object detection models to be bias-free with respect to appearance of the objects?