Utilizing grounded SAM for self-supervised frugal camouflaged human detection

Visually detecting camouflaged objects is a hard problem for both humans and\ncomputer vision algorithms. Strong similarities between object and background\nappearance make the task significantly more challenging than traditional object\ndetection or segmentation tasks. Current state-of-the-art models use either\nconvolutional neural networks or vision transformers as feature extractors.\nThey are trained in a fully supervised manner and thus need a large amount of\nlabeled training data. In this paper, both self-supervised and frugal learning\nmethods are introduced to the task of Camouflaged Object Detection (COD). The\noverall goal is to fine-tune two COD reference methods, namely SINet-V2 and\nHitNet, pre-trained for camouflaged animal detection to the task of camouflaged\nhuman detection. Therefore, we use the public dataset CPD1K that contains\ncamouflaged humans in a forest environment. We create a strong baseline using\nsupervised frugal transfer learning for the fine-tuning task. Then, we analyze\nthree pseudo-labeling approaches to perform the fine-tuning task in a\nself-supervised manner. Our experiments show that we achieve similar\nperformance by pure self-supervision compared to fully supervised frugal\nlearning.\n