-
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
- Python 3
- Pytorch
- Pillow
- Matplotlib
- BeautifulSoup
- tqdm
- The dataset is from Kaggle: https://www.kaggle.com/datasets/andrewmvd/fruit-detection.
- The data is a set of fruit images. There are 4 types of fruits (Banana, Snake fruit, Dragon fruit, Pineapple), each image includes 3-4 objects.
-
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:
