flipkart-grid

Flipkart GRiD – Te[a]ch The Machines | 2019

Level 2

• Vertical Classification using images.
• Develop a model that localizes (bounding box) the object in an image.
  • Given an image model should produce coordinates of the rectangle where in the object lies.

DataSet

• Metadeta file:
  • name-of-image -> (x1, x2, y1, y2)
    • (x1, y1) -> Bottom Left
    • (x2, y2 -> Top Right

Performance metric

• Mean intersection over union of the areas.

Ideas

YOLO v3

• YOLOv3 seems to be the SOTA in object detection.
  • Is there something I am missing?
• Try this First. Seems easy enough.
• Darknet implementation
  • Seems Pretty Straight Forward.

Results

Shortlisted for next Round 😎 😍 😋

1. Yolov3: Only 1 Class

   • Default Everything: 416x416

     • 88.812 Test Score Around 65th Epoch
     • 98 mAP on Validation
     • Training-plots

   • Default + Custom Anchor Points

     • ~85 Percent Test Score
     • 98 mAP on Validation

   • Increased Size of the image (640x640) + custom anchors

     • Took too long
     • 95 mAP on Validation 78

   • Mini

     • Not Promising stuck at 85% Validation

2. Yolov3: 30 Classes (K Mean Clustering)

   • 40 Training Epochs: 74.52 Test Score
   • Training-plots

3. Yolov3: 10 Classes (K Mean Clustering)

   • 50 Training Epochs: 79.9338 %
   • 70 Training Epochs: 81.798 %
   • Training-plots

Level 3

• Same Dataset
• Larger Number of Training and Test Images with update bboxes seems to be the oly difference.

Yolov3

• Pretrained
  • 31 epochs: 87.1057 Test Score