| Summary: | Deep Learning is one of the most approved research trends currently and has brought revolutionary advances in several computer vision applications. Its versatility and robustness motivated us to develop a project which consists of real-time object detection and object retrieval, as part of autonomous wheelchair navigation. Intel's Neural Compute Stick, an edge device was utilized for speeding up inferences. A two-type classification scheme was proposed where firstly, object detection is performed for general class detection and classification followed by Content Based Image Retrieval (CBIR) for object recognition. The aims of this project is to evaluate the effectiveness CBIRsystem combined with deep learning, to tackle several challenges that appear in real-life scenarios, such as camera motion blur, occluded objects, changing size of objects and similar objects and finally, if the use of Transfer Learning and Finetuning for CBIR can provide sufficient results when hardware resources are limited. The object detection part is evaluated with Precision-Recall Curves where an mAPof 75.15% was achieved across three general classes and the CBIR system is assessed with Precision@K measurements and a convincing mean mAP of 89.43% was reached across six speciffic classes.
|
|---|
