Object Detection Using Fast R-CNN
Shriram KV
Bengaluru, Karnataka
Humans have the natural power to identify objects easily on their own, but a machine cannot. Algorithmic descriptions of recognition task has to be implemented on machines to identify an object in an image. Detecting objects and estimating their pose remains as one of the major challenges in computer vision technology. Object recognition is one of the key features in computer vision technology and it is used for identifying a specific object in a digital image or video. The importance of Object recognition algorithm is very high in real-world applications. Compared to image classification, object detection is a more challenging task that requires several complex methods to solve.Some of the applications include Bio-metric recognition, industrial inspection, Robotics, Intelligent Vehicle System, Human-computer interaction, etc. In a retail business, identifying the products of a single manufacturer is difficult. In this paper, Fast R-CNN Algorithm is used to detect the products of a particular manufacturer - [Procter & Gamble]. ...learn more
Project status: Published/In Market
Overview / Usage
An affordable and frugal application for Object Recognition using FAST R-CNN is built.
Methodology / Approach
DataSet Collection: Since the project is all about developing a classifier to detect products of a single manufacturer (i.e Procter & Gamble products ), About 1000 images of products ( Ariel Washing power, Tide
Washing powder, Pantene Shampoo, Head & Shoulders Shampoo, Whisper diaper and Pampers diaper ) has been collected from several supermarket shops. These images were resized to 200kb and less than that. Then the products in these images were labeled accordingly.
Deep Learning with TensorFlow: These images were converted into TFRecords and then we trained these data using object detection Fast R-CNN classifier. This took about 2 hours for training with a good
GPU Machine. After training process, Inference graph and the model were stored separately in the folder.
Fine Tuning: After creating the model, the model was tested in realtime test data and depending on the results, we added some more DataSet for the training to fine tune the model.
Detection and Validation: The final Fine turned model is then tested with the realtime data in a Video Format [.mp4 or .mov], and the results were pretty impressing.
Technologies Used
Tensorflow
Deep Learning