Machine learning models for fine-grained detection and recognition of small objects

In recent years, object detection performance had stagnated. The best performing systems were complex ensembles combining multiple low-level image features with high-level context from object detectors and scene classifiers. As an indispensable and challenging problem in computer vision, small object detection forms the basis of many other computer vision tasks, such as object tracking, instance segmentation, image captioning, action recognition and scene understanding. Although the promising progress of this domain has been achieved recently, there remains a huge gap between the state-of-the-art and human-level performance. In this work, it will be studied the problem of small object detection and classification among a large variety of possible outcomes. Thanks to an hybrid ensemble machine learning approach, it will be possible to better analyse images where the features are embedded in few pixels of information. The fine-graded object are detected using these more robust features than the simple standard segmentation and computer vision approaches. Moreover, it is faced the problem of collecting and labelling the data in a resilient and efficient way thanks to synthetic dataset generation.