I have recently enrolled for a technical research internship cum training program in Deep Learning and Computer Vision. My curiosity for this topic made me search many blogs, tutorials, and LinkedIn profiles. Here is a quick compilation of my research on this topic hope it brings a park in the reader's mind to check out this cool technology and provide a useful application to the world!

Demand for Real Time Image Segmentation Applications

Image segmentation is an aspect of computer vision that deals with segmenting the contents of objects visualized by a computer into different categories for better analysis. The contributions of image segmentation in solving a lot of computer vision problems such as analysis of medical images, background editing, vision in self-driving cars, and analysis of satellite images make it an invaluable field in computer vision. One of the greatest challenges in computer vision is keeping the space between accuracy and speed performance for real-time applications. In the field of computer vision, there is this dilemma of a computer vision solution either being more accurate and slow or less accurate and faster.

PixelLib Library is a library created to allow easy integration of object segmentation in images and videos using a few lines of python code. The previous version of PixelLib uses Tensorflow deep learning as its backend which employs Mask R-CNN to perform instance segmentation. Mask R-CNN is a great object segmentation architecture, but it fails to balance the accuracy and speed performance for real-time applications. PixelLib provides support for the PyTorch backend to perform faster, more accurate segmentation and extraction of objects in images and videos using PointRend segmentation architecture. pointed by Alexander Kirillov et al is used to replace Mask R-CNN for performing instance segmentation of objects. PointRend is an excellent state-of-the-art neural network for implementing object segmentation. It generates accurate segmentation masks and runs at a high inference speed that matches the increasing demand for accurate and real-time computer vision applications. I integrated PixelLib with the python implementation of PointRend by Detectron2 which supports only Linux OS. I made modifications to the original Detectron2 PointRend implementation to support Windows OS. PointRend implementation used for PixelLib supports both Linux and Windows OS.

The images labeled PointRend are obviously better segmentation results than Mask R-CNN.

The images labeled PointRend are obviously better segmentation results than Mask R-CNN.

Download & Installation

Download Python

PixelLib PyTorch supports python version 3.7 and above. Download a compatible python version.

Install PixelLib and its dependencies

Install PyTorch

PixelLib PyTorch version supports these versions of PyTorch(1.6.0,1.7.1,1.8.0 and 1.90). PyTorch 1.7.0 is not supported and do not use any PyTorch version less than 1.6.0. Install a compatible PyTorch version.

Install Pycocotools**

Install PixelLib**

If installed, upgrade to the latest version using:**

Image Segmentation

PixelLib uses five lines of python code for performing object segmentation in images and videos with PointRend model. Download the PointRend model. This is the code for image segmentation.

Line 1-4: PixelLib package was imported and we also imported the class instance segmentation from the module pixellib.torchbackend.instance (importing instance segmentation class from PyTorch support). We created an instance of the class and finally loaded the PointRend model we have downloaded.

Line 5: We called the function segmentImage to perform segmentation of objects in images and added the following parameters to the function:

Image_path: This is the path to the image to be segmented. Show_bbox: This is an optional parameter to show the segmented results with bounding boxes. Output_image_name: This is the name of the saved segmented image.

Sample Image for Segmentation

Image After Segmentation

The checkpoint state_dict contains keys that are not used by the model: proposal_generator.anchor_generator.cell_anchors.{0, 1, 2, 3, 4}

This log above may appear if you are running the segmentation code. It is not an error and the code will work fine.

The segmentation results return a dictionary with values associated with the objects segmented in the image. The results printed will be in the following format:

Detection Threshold**

PixelLib makes it possible to determine the detection threshold of object segmentation.

confidence: This is a new parameter introduced in the load_model function and it is set to 0.3 to threshold the detections by 30%. The default value I set for the detection threshold is 0.5 and it can be increased or decreased using the confidence parameter.

Speed Records**

PixelLib makes it possible to perform real-time object segmentation and added the ability to adjust the inference speed to suit real-time predictions. The default inference speed for processing a single image using Nvidia GPU with 4GB capacity is about 0.26 seconds.

Speed Adjustments** PixelLib supports speed adjustments and there are two types of speed adjustment modes which are fast and rapid modes:

1. Fast Mode

In the load_model function, we added the parameter detection_speed and set the value to fast. The fast mode achieves 0.20 seconds for processing a single image. Full Code for Fast Mode Detection

1. Rapid Mode

In the load_model function, we added the parameter detection_speed and set the value to rapid. The rapid mode achieves 0.15 seconds for processing a single image.

Full Code for Rapid Mode Detection

PointRend Models

There are two types of PointRend models used for object segmentation and they are of resnet50 variant and resnet101 variant. The resnet50 variant is used throughout this article because it is faster and of good accuracy. The resnet101 variant is more accurate but it is slower than the resnet50 variant. According to the official reports of the models on Detectron2, the resnet50 variant achieves 38.3 mAP on COCO and the resnet101 variant achieves 40.1 mAP on COCO.

Speed Records for Resnet101: The default speed for segmentation is 0.5 seconds, the fast mode is 0.3 seconds while the rapid mode is 0.25 seconds.

Code for Resnet101 variant

The code for performing inference with the resnet101 model is the same, except we loaded the PointRend resnet101 model in the load_model function. Download the resnet101 model from here. We added an extra parameter network_backbone in the load_model function and set the value to resnet101.

Note: If you want to achieve high inference speed and good accuracy, use PointRend resnet50 variant, but if you are more concerned about accuracy, use the PointRend resnet101 variant. All these inference reports are based on using Nvidia GPU with 4GB capacity.

Custom Object Detection in Image Segmentation

The PointRend model used is a trained COCO model which supports 80 classes of objects. PixelLib supports custom object detection which makes it possible to filter detections and ensure segmentation of target objects. We can choose out of the 80 classes of objects supported to match our target goal. These are the 80 classes of objects supported:

Code for Segmentation of Target Classes

The function select_target_classes was called to select the target objects to be segmented. The function segment image got a new parameter segment_target_classes to choose from the target classes and filter the detections based on them. We filter the detections to detect an only the person in the image.

Object Extractions in Images

PixelLib makes it possible to extract and analyze objects segmented in an image.

Code for Object Extraction

The code for image segmentation is the same, except we added extra parameters extract_segmented_objects and save_extracted_objects to extract segmented objects and save the extracted objects respectively. Each of the segmented objects will be saved as segmented_object_index e.g segmented_object_1. The objects are saved based on the order in which they are extracted.

This a just a gizt about what I explored in this domain hope to learn more and collaborate on this domain . Keep reading my other blog here .