Recursion in the upsampler does not always terminate

[gfhcs]

Hello,

we are processing the validation set of the REDS 120fps dataset (https://seungjunnah.github.io/Datasets/reds.html). For about half of the sequences there, we encounter a case in which the recursion in upsampling.utils.upsampler.Upsampler._upsample_adaptive does not terminate:

 def _upsample_adaptive(self, I0, I1, t0, t1, num_bisections=-1):
        if num_bisections == 0:
            return [], []

        dt = self.batch_dt = np.full(shape=(1,), fill_value=0.5, dtype=np.float32)
        image, F_0_1, F_1_0 = self.interpolator.interpolate(I0, I1, dt)
        raise Exception()

        if num_bisections < 0:
            flow_mag_0_1_max = ((F_0_1 ** 2).sum(-1) ** .5).max()
            flow_mag_1_0_max = ((F_1_0 ** 2).sum(-1) ** .5).max()
            num_bisections = int(np.ceil(np.log(max([flow_mag_0_1_max, flow_mag_1_0_max]))/np.log(2)))

        left_images, left_timestamps = self._upsample_adaptive(I0, image, t0, (t0+t1)/2, num_bisections=num_bisections-1)
        right_images, right_timestamps = self._upsample_adaptive(image, I1, (t0+t1)/2, t1, num_bisections=num_bisections-1)
        timestamps = left_timestamps + [(t0+t1)/2] + right_timestamps
        images = left_images + [image[0]] + right_images

        return images, timestamps

In the branch if num_bisections < 0: the computation of num_bisections might yield 0, such that recursion becomes infinite (because the recursive calls will hit exactly the same case again) and we either run into Python's recursion limit or run out of memory.

I have a very superficial understanding of this code, but my guess is that adding the following lines as the last instructions inside that if branch fixes the issue:

            if num_bisections == 0:
                return [image[0]], [(t0+t1)/2]

What do you think?

[mandulaj]

We hit the same problem and came to the exact same solution. 👍