Prediction mask resolution when using tiler

[chanthbouala]

Hi everyone,

I've been exploring the various aspects of anomalib library. I'm currently taking a look at tiling the input images in order to feed them to reverse distillation or STFPM model for example. Everything went fine for the training phase. But when I want to predict and collect the predictions with the following lines:

if predictions is not None:
for i, prediction in enumerate(predictions):
image_path = prediction.image_path
anomaly_map = prediction.anomaly_map # Pixel-level anomaly heatmap
pred_label = prediction.pred_label # Image-level label (0: normal, 1: anomalous)
pred_mask = prediction.pred_mask # Pixel-level anomaly mask

I observe the following behavior:

1. The size of image_path is the size of one validation mini-batch
2. Most importantly, the resolution of pred_mask is equal to the tile resolution and not that of the full image. How come? Is there a way to retrieve the full resolution prediction mask knowing that my images are quite big (500x8000)

Thanks for your help.