Özet:

When we have dataset with large number of labelled examples it is easy to perform object detection task but, rare object detection from a few examples is a new problem. Metalearning has been shown to be a promising strategy in the past. However, fine-tuning strategies have received little attention. We discovered that fine-tuning the last layer of detector is a critical task in few-shot object detection. On current benchmarks, such a basic strategy outperforms meta-learning approaches by about 4 to 16 points and sometimes the accuracy is doubled when compared to existing methodologies. However, current benchmarks are frequently unreliable because of the significant variance in the few samples. To generate consistent comparisons, we change the evaluation processes by choosing various sets of training examples. The model has been evaluated on three datasets: COCO, LVIS, and PASCAL VOC. Our fine-tuning approach amalgamated with the Ranking based loss function which can be used for both classification and localization is state-of-the-art.

Anahtar Kelimeler: