Object Detection using YOLOv3

Implementation

Let’s begin by importing the necessary libraries. The OpenCV library is going to be our best friend for this tutorial as it has several helpful functions for manipulating images as well as useful modules such as ‘dnn’.

Since we’ll be using a pre-trained model, we’d have to download certain files. The “weights” file , the “configuration” file, and the “coco-names” file. The weights and the configuration file can be found in this link and the coco-names file can be downloaded/copied from here . There are several pre-trained models available and we would be using the “YOLOv3–416" model. The models are trained on the MS COCO dataset which has 80 classes of objects present in it.

After downloading all the files, its time to create and load our model. As you can see below, the dnn module as several function in-built to aid us in this regard. The names of the objects that our model has been trained to identify is given in the “coco.names” file, which we store in a list called classes. We also retrieve the names of the output layers with the help of the getLayerNames() and getUnconnectedOutLayers() function and store them too in output_layers list.

We now have to pass-in the image through our model. But, we cannot do so directly as our model expects our image to be of a particular shape. This is where the cv2.dnn.blobFromImage() function comes in handy. It helps us to reshape our image while also normalizing them and re-ordering the color channels in proper order.

The image is then given to the model and a forward-pass is performed. The output of which gives us a list of detections. From this list a set of bounding-box co-ordinates for each object detected is obtained as shown below. We use a confidence threshold value to filter out weak detections. The default value I’ve used for confidence threshold is ‘0.5’. All the bounding-box co-ordinates, their class-ids and their corresponding confidence values are stored in lists “boxes”, “class_ids ”and “confidences” respectively.

Now that we have obtained the locations of objects in our image, it’s time to sketch their bounding-box and tag them. The draw_boxes() function does this for us. One problem that we might encounter in our journey is that, the objects, sometimes, may be detected more than once. To avoid such a scenario, we’ll employ Non-Maximum Suppression (aka Non-Maxima Suppression). The default value I’ve used for NMS threshold is ‘0.4’. This is what is performed by the cv2.dnn.NMSBoxes() function down below. We finally display the output image using the cv2.imshow() function.

Now that we have seen all the components required, let us now glue it all together to perform object detection in an image file.

We can perform the same task in videos from files as well as from web-cams as shown below.

The image on the right shows multiple bounding-boxes around the same person. After employing NMS, we obtain the image on the left as output. The duplicate bounding-boxes have been dealt with.