Thread by @ai_fast_track

Many open-world applications require the detection of novel objects.

but state-of-the-art object detection and instance segmentation models are unable to do so.

• It’s because models learn to suppress any unannotated objects by treating them as background

• To address that issue, the authors propose a simple yet surprisingly powerful data augmentation and training scheme they call Learning to Detect Every Thing (LDET)

• To avoid suppressing hidden (unannotated) objects, background objects that are visible but unlabeled, they paste annotated objects on a background image sampled from a small region of the original image (see figure)

• Since training on such synthetically augmented images suffers from domain shift, they decouple the training into two parts:

1- training the region classification and regression head on augmented images, and

2- training the mask heads on original images

• Therefore, a model does not learn to classify hidden objects as background while generalizing well to real images

• LDET leads to significant improvements on many datasets in the open world instance segmentation task

• It outperforms baselines on cross-category generalization on COCO, as well as cross-dataset evaluation on UVO and Cityscapes

⭐ If you are interested in mastering Object Detection (OD), follow me to receive highly curated content right in your feed.

⭐ I started a newsletter, join to master object detection and get a competitive edge in this field

getrevue.co/profile/ai_fas…

📰Paper: Learning to Detect Every Thing in an Open World

abs: arxiv.org/abs/2112.01698
pdf: arxiv.org/pdf/2112.01698…