This project is an extension of the colorization problem solved as part of UMass COMPSCI 689 course I took in Fall 2023.
The data for this problem as of yet is private and currently not hosted anywhere. Will post a drive link for it soon
Project best model
The best model so far can be found here
The above models performs very well on the training set but gives poor performance for the test set. It works well with images with neutral colored objects such as "plane with a blue sky backdrop" but underperforms in many other generic settings.
Project results - this section here is for describing the results of the different models and ablation studies.
| Sr.No | Model | Parameters | Accuracy | Validation Loss | Test Loss | Remarks |
|---|---|---|---|---|---|---|
| 1 | Conv-Net | lr: 0.00008, batch_size: 64, num_epochs: 150 | NA | 0.0684 | 0.0952 | Batch size 64 performed better for training set |
| 2 | Conv-Net | lr: 0.0001, batch_size: 32, num_epochs: 1000 | NA | 0.03804 |
