Generalized Few-Shot Semantic Segmentation: All You Need is Fine-Tuning

Abstract

Generalized few-shot semantic segmentation was introduced to move beyond only evaluating few-shot segmentation models on novel classes to include testing their ability to remember base classes. While all approaches currently are based on meta-learning, they perform poorly and saturate in learning after observing only a few shots. We propose the first fine-tuning solution, and demonstrate that it addresses the saturation problem while achieving stateof-art results on two datasets, PASCAL-5 i and COCO-20 i . We also show it outperforms existing methods whether finetuning multiple final layers or only the final layer. Finally, we present a triplet loss regularization that shows how to redistribute the balance of performance between novel and base categories so that there is a smaller gap between them.

Robust 3D Object Detection in Cold Weather Conditions

“…Additionally, they apply other FSS methods developed for imagery to 3D point clouds. In particular, they study the use of knowledge distillation combined with fine-tuning [15] and fine-tuning combined with contrastive learning [16]. In this work, we use [14], [15], [16] as FSS baselines for our method.…”

“…FSS Baselines. As baselines for FSS, we use three stateof-the-art approaches for outdoor LiDAR point clouds (FS-SAD [14], GFSS [16] and LwF [15]). For FSSAD [14], we train ϕ S0 using the cross-entropy and Lovász-softmax functions described by the authors.…”

“…For FSSAD [14], we train ϕ S0 using the cross-entropy and Lovász-softmax functions described by the authors. For GFSS [16], we train ϕ S0 using the cross-entropy and triplet loss regularization. For the latter, we use a maximum of 50 anchors per training step and set the margin parameter to 1.…”