Object Detection Models with Limited or Small Labeled Datasets

Deep Learning made a substantial improvement in the results of different computer vision tasks. However, the use of deep learning models comes with a significant cost: the requirement for substantial amounts of labeled data, also known as annotated data.

Different techniques have been developed in order to train deep learning models with a limited number of labeled data samples. In image classification, this problem - of limited labeled data - has been addressed with great success, but in object detection, the problem is more complex given its nature where two learning tasks are performed: the localization of the boxes - a regression step - and the classification of each box, which makes object detection more complicated.

The main concern in such models revolves around the prudent utilization of labeled samples to yield optimal results when performing inferences. Addressing this issue is a hard challenge and is the core of this dissertation, which delves into diverse strategies for harnessing labeled data to enhance model performance.

We investigated three methods to enhance performance when limited labeled data is available. First, we looked at how to preprocess the available data using multispectral channels, metadata, and data augmentation to increase performance. Secondly, we studied semi-supervised learning in object detection by refining the pseudo-labels using additional models before and after the non-maximum suppression (NMS) method. Lastly, we explored the concept of direct inference object detection, using a foundational visual model (VFM) and few-shot models (from the labeled dataset) to make direct inferences without any training steps.

In all cases, we show how there is a significant improvement, shedding light on potential directions for future research in this domain.