Python Runner on Linux: Squash not Crop

MMarcial · February 7, 2023, 5:41pm

Is there a way to squash the image instead of cropping the image. get_features_from_image() crops the image. The Data Input Block in EI Studio allows squashing.

with ImageImpulseRunner("demo-linux-x86_64-v2023.eim") as runner:
   features, cropped = runner.get_features_from_image(img)

louis · February 10, 2023, 9:21am

Hello @MMarcial,

Good question, we’ve been speaking about that in the past but I’m not sure it’s been implemented.
Let me have a deeper look at the code, if it’s not I’ll create an issue.

Best,

Louis

louis · February 10, 2023, 9:34am

Hello @MMarcial,

So indeed, it is missing from the function:

github.com

edgeimpulse/linux-sdk-python/blob/c440f40daef2e3f8497d7e67e31c5010dfd50ca1/edge_impulse_linux/image.py#L73

      
        
                while not self.closed:
                    success, img = self.videoCapture.read()
                    if success:
                        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
                        features, cropped = self.get_features_from_image(img)
            
            
            res = self.classify(features)
                        yield res, cropped
            
            
# This expects images in RGB format (not BGR)
            def get_features_from_image(self, img, crop_direction_x='center', crop_direction_y='center'):
                features = []
            
            
    EI_CLASSIFIER_INPUT_WIDTH = self.dim[0]
                EI_CLASSIFIER_INPUT_HEIGHT = self.dim[1]
            
            
    in_frame_cols = img.shape[1]
                in_frame_rows = img.shape[0]
            
            
    factor_w = EI_CLASSIFIER_INPUT_WIDTH / in_frame_cols
                factor_h = EI_CLASSIFIER_INPUT_HEIGHT / in_frame_rows

I have not tried the following modification but that’s how I’d do it:

Replace

resized = cv2.resize(img, resize_size, interpolation = cv2.INTER_AREA)

by

resized = cv2.resize(img, (EI_CLASSIFIER_INPUT_WIDTH, IEI_CLASSIFIER_INPUT_HEIGHT), interpolation = cv2.INTER_AREA)

Feel free to use other resize method but I’ve figured over time that INTER_AREA is usually the one giving the best results/performance.

Then get rid of the crop conditions below (between lines 93 and 111).

Finally just use:

cropped = resized

Let me know if that works, I’ll create an internal issue so we can natively support that.

Best,

Louis