In this competition, we’re challenged to develop an algorithm that automatically removes the photo studio background. This will allow Carvana to superimpose cars on a variety of backgrounds. You’ll be analyzing a dataset of photos, covering different vehicles with a wide variety of year, make, and model combinations.
challenge : Link
Input Image Resolution :
- width : 1918
- height: 1280
- 5088 training images
- 100064 test images
The metric used to score this competition requires that your submissions are in run-length encoded format.
I have trained 3 different models on image size = (320, 480):
- UNet architecture with pretrained MobileNetV2 encoder.
- DeeplabV3p with MobileNetV2 architecture with pretrained cityscapes weights
- DeeplabV3p with MobileNetV2 architecture with pretrained MobilenetV2 encoder
Key points
- For training and generating test results i have used efficient data pipeline tf.data
- Adam Optimizer
- dice_loss + bce_jaccard_loss from segmentation-models API
- Batch size = 8
I have recieved these results on training each model for just 10 epochs.
| Model | backbone | score | Remark |
|---|---|---|---|
| Unet | mobilenetV2 | 0.92799 | with pretrained encoder |
| DeeplabV3p | mobilenetV2 | 0.99217 | with pretrained cityscapes weights |
| DeeplabV3p (custom) | mobilenetV2 | 0.99217 | with pretrained encoder only |