Im trying to interpret the YOLOv4 output using TensorFlowLite on the iOS. I have read a little about the concept of cells and anchors used in the output tensor, and I would like to implement the parser for it. Output of my model consists of 2 tensors:
float32[1,13,13,255]
float32[1,26,26,255]
First thing Im wondering about is what is in the second output? From what I have read the first one should contain all the information. Is the second one just a more detailed result or something else?
Ok, but let's start with the first output. I have written a simple function that should extract all cells and then all anchors data from these cells. This is how it looks:
let output0 = try localModel.output(at: 0)
guard let output0Floats = [Float](bytes: output0.data) else { return nil }
let numberOfCells = 13
let numberOfAnchors = 3
let numberOfClasses = 80
let anchorSize = (numberOfClasses + 5)
func cellAt(x: Int, y: Int) -> [Float] {
let cellSize = anchorSize * numberOfAnchors
let position = (y * numberOfCells + x) * cellSize
return [Float](output0Floats[position..<position + cellSize])
}
func anchors(in cell: [Float]) -> [[Float]] {
(0..<numberOfAnchors).map { [Float](cell[$0 * anchorSize..<$0 * anchorSize + anchorSize]) }
}
for y in 0..<numberOfCells {
for x in 0..<numberOfCells {
let cell = cellAt(x: x, y: y)
print("Cell: \(x),\(y) contains anchors:")
print(anchors(in: cell))
}
}
...
private extension Array {
init?(bytes: Data) {
guard bytes.count % MemoryLayout<Element>.stride == 0 else { return nil }
self = bytes.withUnsafeBytes { .init($0.bindMemory(to: Element.self)) }
}
}
And this is the example result I'm getting for a single cell:
Cell: 7,12 contains anchors:
[[0.693655, -1.1966848, -0.007975042, -0.3327814, -9.583811, 0.3976263, -6.0192285, -6.329881, -5.8644676, -10.2914715, -9.632221, -8.071436, -6.399925, -5.240812, -8.791572, -5.6437893, -9.8603115, -10.492198, -1.9372412, -7.0640965, -2.6936512, -5.112247, -7.131972, -7.1825066, -7.4413238, -10.401382, -7.5643044, -8.608834, -8.239082, -6.799241, -8.035741, -5.7502255, -8.881622, -7.3571744, -9.315964, -7.925786, -7.7857537, -4.8930154, -8.529579, -7.633353, -8.817726, -7.47082, -8.291334, -4.683982, -4.170734, -6.193165, -7.8437185, -9.854808, -9.490823, -8.272433, -8.434413, -7.765057, -7.149798, -11.194118, -6.5116143, -11.112444, -9.999684, -10.689343, -9.942104, -9.520727, -7.440444, -2.531265, -3.7234814, -7.5839844, -4.550161, -3.031804, -4.616852, -8.832014, -6.0279136, -9.482858, -6.750441, -8.450063, -10.222086, -7.6301804, -7.559189, -10.234117, -6.999834, -7.1350074, -5.308107, -6.2450233, -8.8833885, -9.381562, -3.8812854, -8.868278, -9.988986], [0.4351927, -1.3958519, 0.46428338, -0.39240548, -8.170114, 0.7084342, -7.709829, -5.9856057, -6.808081, -10.644019, -9.912677, -7.3293757, -7.548369, -5.533275, -10.072926, -7.316476, -9.945337, -11.118561, -3.2463353, -10.561513, -5.067392, -7.312641, -8.729989, -9.5539055, -7.58917, -9.886164, -6.5404315, -8.553915, -9.023286, -9.580754, -6.7592535, -8.380334, -8.182065, -7.2239976, -9.276712, -7.5086412, -7.2454534, -7.139829, -8.614485, -7.8158274, -9.850543, -9.123642, -6.8081083, -6.936388, -7.997142, -8.845028, -11.322939, -10.713314, -9.629859, -10.820017, -10.480835, -9.071951, -7.9244685, -12.562474, -7.1654305, -13.456438, -10.116255, -12.255847, -11.530319, -10.3949375, -10.665162, -5.6975913, -4.050809, -10.665826, -2.638548, -3.5531735, -7.0320325, -10.047072, -7.678191, -10.290669, -7.438999, -7.531754, -9.817409, -8.428637, -9.502961, -10.955662, -8.6340065, -5.0168147, -8.593948, -9.412493, -10.816083, -10.903126, -8.81499, -10.449745, -9.069517], [0.025469145, -1.7808459, -0.18256505, -0.70104045, -10.450736, -0.67288893, -5.771856, -5.448979, -6.4159226, -8.777289, -7.960696, -5.3555217, -4.798117, -2.8378687, -7.9489646, -8.255625, -8.968552, -8.036578, -2.46956, -8.458385, -4.8979797, -6.5746903, -7.2408285, -8.574903, -6.8356185, -6.4320874, -6.037178, -7.56021, -7.275848, -8.808907, -3.9019513, -8.835796, -6.360187, -6.5461373, -7.1117754, -6.6027184, -7.280362, -7.1671834, -7.292713, -7.1488175, -7.1398635, -8.180893, -5.797153, -6.3417816, -6.9332256, -8.371075, -9.2042055, -8.602686, -8.072069, -8.1690035, -8.0164175, -6.61691, -6.3536263, -9.318304, -4.5542707, -10.049933, -7.8087454, -9.497473, -9.07455, -8.406244, -7.078502, -5.5775504, -2.3586287, -8.409487, -1.6716739, -3.8225765, -6.9020715, -6.6682305, -5.784493, -8.40492, -7.2747784, -6.392035, -6.4958863, -7.629692, -7.4995623, -8.4432125, -6.7565637, -3.113231, -7.3596015, -8.573539, -8.829562, -8.523581, -8.571439, -8.087017, -7.958835]]
So single anchor looks like this:
[0.693655, -1.1966848, -0.007975042, -0.3327814, -9.583811, 0.3976263, -6.0192285, -6.329881, -5.8644676, -10.2914715, -9.632221, -8.071436, -6.399925, -5.240812, -8.791572, -5.6437893, -9.8603115, -10.492198, -1.9372412, -7.0640965, -2.6936512, -5.112247, -7.131972, -7.1825066, -7.4413238, -10.401382, -7.5643044, -8.608834, -8.239082, -6.799241, -8.035741, -5.7502255, -8.881622, -7.3571744, -9.315964, -7.925786, -7.7857537, -4.8930154, -8.529579, -7.633353, -8.817726, -7.47082, -8.291334, -4.683982, -4.170734, -6.193165, -7.8437185, -9.854808, -9.490823, -8.272433, -8.434413, -7.765057, -7.149798, -11.194118, -6.5116143, -11.112444, -9.999684, -10.689343, -9.942104, -9.520727, -7.440444, -2.531265, -3.7234814, -7.5839844, -4.550161, -3.031804, -4.616852, -8.832014, -6.0279136, -9.482858, -6.750441, -8.450063, -10.222086, -7.6301804, -7.559189, -10.234117, -6.999834, -7.1350074, -5.308107, -6.2450233, -8.8833885, -9.381562, -3.8812854, -8.868278, -9.988986]
Now I can't understand these numbers. From what I read, first 5 numbers should be:
Confidence, BBoxX, BBoxY, BBoxWidth, BBoxHeight
and the rest of the values are probabilities of each class in a labelMap.
But these numbers look totally incorrect to me. Shouldn't Confidence be between 0 and 1? And probabilities shouldn't be between 0 a 1? What can I be doing wrong that I'm getting these results? The code I'm using before parsing these results is well tested with other types of tflite files, and it should be fine. Can this be due to the incorrect imageMean and imageStd used in the input pixel buffer preparing? I'm not sure which values were used to build this model, so Im using 127.5 for both of these values.