README.md

Evaluation Results

There are likely to be some hidden hyperparameter choices that are not replicated exactly in these experiments.

In the paper, every dataset in split in half into train / test sets. Since the exact split is unknown, all my experiments simply evaluate on the entire dataset.

Splice Detection

Average precision scores are reported.

	Results from Paper	This Implementation
RT	0.55	0.54
Columbia	0.98	0.95

Splice Localization

Class-balanced IOU (cIOU) scores are reported.

I resize all the ground-truth and prediction maps into the same size in order to compute the optimal threshold and corresponding IoU score for each image in a vectorized and efficient manner.

	Results from Paper	This Implementation
RT	0.54	0.54
Columbia	0.85	0.67
In-the-Wild	0.58	0.64
Hays	0.65	0.54

Qualitative Results

Realistic Tampering

Columbia

In-the-Wild

Scene Completion (Hays)