This repository builds an end-to-end multi-class, multi-object image detector using RCNN which is a popular algorithm for object detection.

Paper: Rich feature hierarchies for accurate object detection and semantic segmentation

The workflow is follow strickly by the pipeline below:

Specifically, in my repository, RCNN algorithm is applied as below:

• Step 1: The Selective Search algorithm is applied on each image to find the proposed ROIs (Region of Interest). After that, these regions are divided into 5 classes ('nothing', 'pineapple', 'snake fruit', 'dragon fruit', 'banana'). The coordinates of bounding box and label of each region are saved.
• Step 2: A fine-tuned EfficientNet model is used to train for image classification problem on above dataset.
• Step 3: At the beginning of inference phase, the Selective Search is applied to find the proposed boxes on infer image.
• Step 4: Next, fine-tuned model trained above is used to predict class of each box.
• Step 5: Apply Non-Maximum Suppression (NMS) algorithm to remove redundant boxes.
• Step 6: Return final object detection results.

A sample result is described below: