Different results using DefaultPredictor vs model(inputs)

[nrasiwas]

Getting two different results on the same image using DefaultPredictor vs model(inputs). Here are the two different outputs

Code snippets:

im = cv2.imread("...")
outputs = predictor(im)
print(outputs["instances"].pred_classes)
print(outputs["instances"].pred_boxes)

Using model:

from detectron2.modeling import build_model
model = build_model(cfg)  # returns a torch.nn.Module

from detectron2.checkpoint import DetectionCheckpointer
DetectionCheckpointer(model).load(cfg.MODEL.WEIGHTS)  # load a file, usually from cfg.MODEL.WEIGHTS

image = torch.as_tensor(im.transpose(2, 0, 1))
inputs = [{'image': image}]

model.eval()
with torch.no_grad():
    outputs = model(inputs)

print(outputs[0]["instances"].pred_classes)
print(outputs[0]["instances"].pred_boxes)

[ljtruong]

The difference is preprocessing.

This is a reference to the defaultpredictor which shows the preprocessing step before applying model(inputs).

[nrasiwas]

Thanks. I was able to replicate the results, although I have a confusion now. My original image was 500x458 and the transform upsamples it to 873x800. Is this really helpful? Processing on the larger image is 3x slower (because it has 3x more pixels), but there is no extra information.

ResizeTransform(h=500, w=458, new_h=873, new_w=800, interp=2)

Do you think I will loose much by directly operating on 500x458 size images?

[ljtruong]

you can change the default input sizes using cfg.INPUT.MIN_SIZE_TRAIN which defaults to 800 as the smallest size.
The aim of resizing in this manner is about maintaining aspect ratio.

in your case training on smaller image sizes wouldn't lose anything as the images are smaller than the default resize.

[nrasiwas]

I am actually using pertained models from the model zoo. Do you think if they were trained on 800 as minimum size but I have 500px images, is better to upscale or its fine to just feed in 500px itself?

…

On Thu, May 27, 2021, 5:34 PM Leon Truong ***@***.***> wrote: you can change the default input sizes using cfg.INPUT.MIN_SIZE_TRAIN which defaults to 800 as the smallest size. The aim of resizing in this manner is about maintaining aspect ratio. in your case training on smaller image sizes wouldn't lose anything as the images are smaller than the default resize. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#3097 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACCG4JONH4UMF2XWL53ORPTTPYYLNANCNFSM45TUG6PA> .