ROMAn-DETR: We came, We sawed, We built a Robust Orientation-aware Model with Anisotropic attention

Abstract

Objects in industrial sawmill environments tend to be in complex lighting scenarios, paired with them being highly geometrically anisotropic (long and thin) while lying close to several other similar objects. This makes them unsuitable for axis-aligned bounding boxes, which is a problem for segmentation and detection models. To solve this we introduce our model ROMAn-DETR, an instance segmentation model made to handle anisotropic objects while also improving robustness for outof- distribution data. By combining geometric priors for these anisotropic objects, together with a new auxiliary orientation module capable of predicting the axis of the objects’ search space, the search space can be reduced for more efficient per-query sampling. Through extensive ablations we show that ROMAn-DETR both converges faster and also outperforms an equivalent state-of-the-art RF-DETR model in mAP by up to 10.3% on our out-of-distribution anisotropic dataset. This paper also evaluates potential alternative approaches using a smaller head together with a DINO backbone which showed that while DINO is robust, it by itself not enough for instance segmentation tasks.