Memory leak

[Neural Ant]

Hi, thank you for your efforts.
I would like to use your library, but there is a memory leak somewher.
If I instanciate only one UltraFace and get only one Mat Cv2.ImRead(imageFile) and than use ultraFace.Detect in a loop, it leaks memory very fast.
Can you help me with that?

[Neural Ant]

Hi Takuya,

I have found it! It's in the GenerateBBox method:

` private void GenerateBBox(ICollection boundingBoxCollection, Mat scores, Mat boxes, float scoreThreshold, int numAnchors)
{
for (var i = 0; i < numAnchors; i++)
if (scores.Channel(0)[i * 2 + 1] > scoreThreshold)
{
var rects = new FaceInfo();
var xCenter = boxes.Channel(0)[i * 4] * CenterVariance * this._Priors[i][2] + this._Priors[i][0];
var yCenter = boxes.Channel(0)[i * 4 + 1] * CenterVariance * this._Priors[i][3] + this._Priors[i][1];
var w = Math.Exp(boxes.Channel(0)[i * 4 + 2] * SizeVariance) * this._Priors[i][2];
var h = Math.Exp(boxes.Channel(0)[i * 4 + 3] * SizeVariance) * this._Priors[i][3];

                rects.X1 = Clip(xCenter - w / 2.0, 1) * this._ImageW;
                rects.Y1 = Clip(yCenter - h / 2.0, 1) * this._ImageH;
                rects.X2 = Clip(xCenter + w / 2.0, 1) * this._ImageW;
                rects.Y2 = Clip(yCenter + h / 2.0, 1) * this._ImageH;
                rects.Score = Clip(scores.Channel(0)[i * 2 + 1], 1);

                boundingBoxCollection.Add(rects);
            }
    }`

Can you chek if my understanding is correct, and my soltuion is ok?
`private void GenerateBBox(ICollection boundingBoxCollection, Mat scores, Mat boxes, float scoreThreshold, int numAnchors)
{

        var scoresChannelZero = scores.Channel(0);
        var boxesChannelZero = boxes.Channel(0);

        for (var i = 0; i < numAnchors; i++)
            if (scoresChannelZero[i * 2 + 1] > scoreThreshold)
            {
                var rects = new FaceInfo();
                var xCenter = boxesChannelZero[i * 4] * CenterVariance * this._Priors[i][2] + this._Priors[i][0];
                var yCenter = boxesChannelZero[i * 4 + 1] * CenterVariance * this._Priors[i][3] + this._Priors[i][1];
                var w = Math.Exp(boxesChannelZero[i * 4 + 2] * SizeVariance) * this._Priors[i][2];
                var h = Math.Exp(boxesChannelZero[i * 4 + 3] * SizeVariance) * this._Priors[i][3];

                rects.X1 = Clip(xCenter - w / 2.0, 1) * this._ImageW;
                rects.Y1 = Clip(yCenter - h / 2.0, 1) * this._ImageH;
                rects.X2 = Clip(xCenter + w / 2.0, 1) * this._ImageW;
                rects.Y2 = Clip(yCenter + h / 2.0, 1) * this._ImageH;
                rects.Score = Clip(scoresChannelZero[i * 2 + 1], 1);

                boundingBoxCollection.Add(rects);
            }

        scoresChannelZero.Dispose();
        boxesChannelZero.Dispose();
    }`

We coul use using blocks, I just wanted to be explicit.

[Takuya Takeuchi]

Thank you for kindly support and good investigation!!
Exactly. NccDotNet.Mat.Channel(int) returns new Mat object.
So result of Channel method should be disposed.

OK, let me test fixed code.

[Takuya Takeuchi]

@NeuralAnt
OK. Fixed.
Could you check the latest develop branch?
You can test by only compiling UltraFaceDotNet without NcnnDotNet module.

After your confirmation, I will publish new nuget package 1.0.0.1.