S13

Session 13 - Object Detection with YOLO v3

Part 1 - YOLO v3 with OpenCV

Detecting objects in an image where there is a person and an object present in the image.
Note: The object should be present in the COCO classes.

Result

Part 2 - YOLO v3 with PyTorch

Object detection with Yolo v3 using transfer learning on a class that doesn't belong to COCO dataset.

Class selected: LEGO Batman

Result

Click on the video below to play

Parameters and Hyperparameters

• Number of images: 500
• Batch size: 10
• Epochs: 300

Dataset Preparation

For using the LEGO Batman dataset, follow the instructions mentioned here. To run the model on custom dataset, follow the steps below

Train Data

• Annotating the images
  • Clone the annotation tool from this link.
  • Follow the steps mentioned in the README of the tool specified above.
  • Annotate atleast 500 images with the tool.
• Creating dataset directory
  • Place the annotated images here.
  • Place the labels here.

Test Data

• Download a short-duration video containing the class used during training.
• Extract frames from the video into the test directory
  ffmpeg -i video.mp4 data/test/img%3d.jpg
• Extract audio from the video (this audio will be required later)
  ffmpeg -i video.mp4 -f mp3 -ab 192000 -vn audio.mp3

Downloading Pre-Trained Weights

Download the file named yolov3-spp-ultralytics.pt from this link and place it in this directory.

Inference on a Video

• Combine the images from the output directory to form a video
  ffmpeg -framerate 24 -i YoloV3/output/img%3d.jpg -r 24 -y out_video.mp4
• Combine the audio file extracted earlier with the output video to produce final output
  ffmpeg -i out_video.mp4 -i audio.mp3 -shortest result.mp4

Results

After running the algorithm for 300 epochs, the result is pretty amazing!

Group Members

• Rakhee (Canvas ID: 25180625)
• Shantanu Acharya (Canvas ID: 25180630)