Problems with Normalize node

[hgmarques]

Hi guys,

I've noticed that the Normalize node not always converts the image pixels to values in the range of [0,1] (or at least the IplImageVisualizer on the Normalize node shows values outside of the [0,1] range, image attached). I don't really understand why since the source code seems to be fine.

I attached a simple Bonsai example to reproduce my observations.

Any help would be welcome. Thanks you in advance.

Cheers,
hugo

NormalizeBug.zip

[bruno-f-cruz]

It looks funky. Looking at the function bonsai implements:

https://github.com/bonsai-rx/bonsai/blob/404b092b501fbd9a0db5ccce76bf695623e6ba7c/Bonsai.Vision/Normalize.cs#LL34C17-L34C32

                var range = max - min;
                var scale = range > 0 ? 1.0 / range : 0;
                var shift = range > 0 ? -min : 0;
                CV.ConvertScale(input, output, scale, shift);

shift is always in raw units.
However, the opencv function follows the canonical algebraic order of operation and applies the scaling first and the shift second.

It seems the easiest way to correct this would be to normalize the shift by the range:

var shift = range > 0 ? -min*scale : 0;

Here's a temporary fix that you can place in your extensions folder.

using OpenCV.Net;
using System;
using System.ComponentModel;
using System.Linq;
using System.Reactive.Linq;

namespace Bonsai.Vision
{
    /// <summary>
    /// Represents an operator that normalizes the range of values for each image
    /// in the sequence to be between zero and one.
    /// </summary>
    [Description("Normalizes the range of values for each image in the sequence to be between zero and one.")]
    public class NormalizeFix : Transform<IplImage, IplImage>
    {
        /// <summary>
        /// Normalizes the range of values for each image in an observable sequence
        /// to be between zero and one.
        /// </summary>
        /// <param name="source">The sequence of images to normalize.</param>
        /// <returns>
        /// The sequence of normalized images.
        /// </returns>
        public override IObservable<IplImage> Process(IObservable<IplImage> source)
        {
            return source.Select(input =>
            {
                var output = new IplImage(input.Size, IplDepth.F32, input.Channels);
                double min;
                double max;
                Point minLoc;
                Point maxLoc;
                CV.MinMaxLoc(input, out min, out max, out minLoc, out maxLoc);
                var range = max - min;
                var scale = range > 0 ? 1.0 / range : 0;
                var shift = range > 0 ? -min*scale : 0;
                CV.ConvertScale(input, output, scale, shift);
                return output;
            });
        }
    }
}