Accuracy with insightface

[hashimkhanzada]

I have a web app connected to a cctv camera using human (insightface).
I've been having some problems with accuracy even though I've got a large enough dataset for each person (about 100 images for each) and also changed the default model to insightface-efficientnet.

I ran a basic test using some sample images, and the results are a bit odd. Not sure if the crop factor is the main cause, or if there are others.

My first image (the one with similarity 100%) is the one I'm using to compare to the rest. The images get ordered by distance once the detections are done.
The expectation is that it has all my images shown first with a relatively high similarity (low distance) and has the rest of them much lower.

match options for all models (this was just to test it out):
{ order: 2, multiplier: 1, min: 0, max: 1 }

When running the test using the default model (face-res), The images were in the correct order. The problem with this is the similarity is quite close between all of them. Increasing the multiplier etc widens the similarity gap which technically makes the celebrity faces go below the match threshold (50%), but it also makes my faces closer to the 50% mark, almost making them a false negative.

Default embedding (description: true)

Tuned match options ({ order: 2, multiplier: 25, min: 0.2, max: 0.8 }):

Running the same test on insightface, it doesn't have the correct order anymore. I'm assuming the crop factor could be a variable to tune, but I couldn't find a way to update it. With the default model it can be tuned using 'description: { enabled: true, crop: 0.8 }' but doing the same for insightface ('insightface: { enabled: true, crop: 0.8, modelPath:'.....' }') doesn't seem to do anything.

Insightface efficientnet (https://vladmandic.github.io/insightface/models/insightface-efficientnet-b0.json)

Insightface mobilenet also has the incorrect order, but slightly different ('https://vladmandic.github.io/insightface/models/insightface-mobilenet-swish.json').

I'm not quite sure why the accuracy is off, apart from the crop factor I tried messing around with almost all the config options but it didn't seem to make much of a difference. I tried the image comparisons on some online demo websites.

https://3divi.ai/demo

The 1:1 Verification demo gave a 100/100 score on all my images (except for the zoomed in one as it couldn't find a face in it). It gave a similarity of less than 5/100 for all the celebrity images.

Another web demo, same results. Got a 99% match on all my images, 0% match on the others.

https://faceapi.regulaforensics.com

Can you please advise as to what the main factors could be for these accuracy ratings

Thanks

My human config:

config: any = {
    // user configuration for human, used to fine-tune behavior
    cacheSensitivity: 0,
    modelBasePath: '../../../../assets/face-models/',
    filter: { enabled: true, equalization: true },
    debug: true,
    backend: 'humangl',
    face: {
      enabled: true,
      detector: {
        rotation: true,
        return: true,
        mask: false,
        maxDetected: 10,
        skipFrames: 0,
        skipTime: 0,
      }, // return tensor is used to get detected face image
      mesh: { enabled: true },
      description: {
        enabled: false,
      }, // default model for face descriptor extraction is faceres
      insightface: {
        enabled: true,
        modelPath:
          '../../../../assets/face-models/insightface-efficientnet-b0.json',
      },
      iris: { enabled: true }, // needed to determine gaze direction
      emotion: { enabled: false }, // not needed
      antispoof: { enabled: false }, // enable optional antispoof module
      liveness: { enabled: false }, // enable optional liveness module
    },
    body: { enabled: false },
    hand: { enabled: false },
    object: { enabled: false },
    gesture: { enabled: false }, // parses face and iris gestures
  };