# Inference Accuracy

Qualcomm® Neural Processing SDK inference classification precision is measured against
several popular public models. Based on our measurement, the
accuracy score do not vary per chipset.

Qualcomm classification precision metrics

The following clasification precision scores are computed by
comparing the Qualcomm® Neural Processing SDK inference result with the ground truth:

- mAP: mean average presion
- Top-1 error rate: chance the highest-probability predicted
class is not the real class
- Top-5 error rate: chance the real class is not contained in
the 5 classed with highest probablity

Mean Average Precision Calculation

mAP (mean Average Precision) is the [Average
Precision](https://en.wikipedia.org/wiki/Evaluation_measures_%28information_retrieval%29#Average_precision)
across all categories. Each AveP (Average Precision) is
calculated by:

![../images/AP_formula.png](data:image/png;base64,UklGRg4GAABXRUJQVlA4TAEGAAAv+UAPAIegIG0DFs7An4K0DVg4A38K0jZg4Qz8zX8A2NY2bf2XxDZt2WabbZJIQhJJtv0HWrZtx22ji4YAiO6d+U+XACQ5SvxVtSL6PwH8R3HVaae5YEnhh/VYj+6r/0E91qP76l+OqMSzO3XLXWk0/awUBuc95D8mhcF5D/nLTaf2oKhDU+cTVgKQ9QxtUbwBRGmDufuHTSk+ijeAIvUV5m7gvvq34pazx3zkADm+bbjCMwhm0ACmGssQd1TNn8VQAaYayyz2VRPoIX8rLsXHMmYI5V3mblYRaCpA8qxdPsDljcWxqukjXR1InvUVDnAZwNz9rczp3qmB+m3vSYG1yuIGXFpNtZPL74Kbj6r6kVsALm18Prn8g6r5pWjSWHVNCK7y1qm66upA1IShsqqaNKksimyFhWBQVVlbVKZ5VWAmpzCBKwBDZTVVmU5xUWQPXP5WXLpWn701V0UwLt1AV18lD6bEpm+wEKyrsI0ju1xNE5rLNlQAl4GuvqqaEIJt+uLyX2sq/4RLrKMkXY0zn7Bw8wbMBZc4za7wNJfBdMNUO0seUjJeKLJvFQM/McaNS+x3pruHyluochwHzzI7my0Eg6p5Fq4ZM9uT/qWKm7uo/IaZytlQPRgqiybJcWIHwwWXTkxl5SUXbwOSBxgqC5OkcWLfratxqPaS5eheuDUOcGmRZVV9d3n2FoINpYOuvlJm668HLi2a2lTZXZ6vZj6zDB1MNqVyRCcuLq0ucZr8IgSm8s7lnYVgUHXviuxoZDCVRfKLLMOHncs7+0pXYDnVoXkIl9G1LsDm1nxMJzuZqg9lCN7SfFTNjYVgAFV1Ez3r6CfVD+hqPQNT9REDZM27Pqrm6vJ849vNVXrkNHLn5dVUfUhyJ2Rv0NWgOT8AzBV2yVjeO6WNRSlNwIKyAWRvgDJM7yqAucLa5W80dNohSIl3kfzj9RQG5yNk/tQUBucjZNZV80eYi4umpy8fi0cTU8/u0bUuR9Pdb6HHdnRfnT+3x3Z0X521ucKPLFqR1bGk+1OvNjHU6uR8umvBcOUtf8ERMj9yuux3gCn+rBzhcjfnepYHVuLf6L46P/O6i1YxLq6f9U/fPTsVoKj+mkLddXVm0cVvuUYoGo8iSeFmX3T6GzHXH/1xeX6m/qE/VLRcuPRDfotTHeiLofuVotNfyHWxuIFb/ZXfbVd7FGWwqPaL6k5hwi25TpQ0fk9/dolwx3dZcnIfmFI861L5t0oJyrssXNbCBxgqZzT1f6tY4CpvKpp8tqu/UMW39OXv4BvE+z3TBz58ixdT+BZF/W+gfd9Y8XLlUfTcZQVGVAJmcgoTuMLBvKTs88OKlysPK15KVO86ED1U76Yl+WlJfj6sePmGRWWaVwHT0wPZSXF8LYbANNlaCJOiDjBU2efaVBKhQnPZhgrg8m64a3KpARbCpKiDhdCJsSZTAVyiJlNOsyunOVSA6eM0Hwkjq95cAoiBZ3GDqv69uoMu9kkTUHncmgfQdPE0l8F0w1TbBW9QNIGkCahA1oBSYOqGoQqYvIHchEUMQLiA7DLkhUuLGIFg36tEqGFnShwkx3HWXGRns4VgUDU3TRVIHjAlVqbEumpA1QCaKnQV6GrQ1K1fbgDBAzECQ3WRXOeLF23jqqkCXffDx7MQWXrJxduA5NkmGeAT0FSBrpuqtkkOSA4gy6DIoMggSYp5AqYMqAC35sKCkn0viAWuwraoA1kTmCpnKitltv7aRQd03UBRB7ImRbbxF5iLjxiAywMxADGw7WrQ1IHk2BYXvlmoEF8xFwGa+lFXOxoZTOXAg0X1nbkIRaxNBZIrD2XAXYDyzi6gyKAIIMaFA5r6F1MH13dTiRncfGTZUdFcRT+pfkBX63mVNGasmndlKjGDm9B1YyXDVCZVlZ7pqjB1P8rIFHV6qED0QJKlSfTWI1MFS86+19QEHVC9XJo8g+c4etYWpTQBC8q2sqA4pnM3UL1cmgDFyRcDsnwjKhlFE6o6EJUMshQ7gBLQ5BtUp6szk5di56+sdParbmq/NQuBX3rTZb+1/9sIAA==)

Where:

- k is the rank in the sequence of retrieved documents
- n is the number of retrieved documents
- P(k) is the precision at cut-off k in the list. The
precision is calculated by tp/(tp+fp) where tp is true
positives, fp is false positives.
- rel(k) is an indicator function equaling 1 if the item at
rank k is a relevant document, zero otherwise.
- the precision score is zero if no relevant documents get
retrieved.

python code example to calculate AP:

for j in range(len(img_sorted)):
        if img_sorted[j] in anno_imgs:
            count += 1.0
            AP += count/rank
        rank += 1.0
    if (count == 0):
        AP = 0
    else:
        AP = AP/count
    Copy to clipboard

Last Published: Oct 02, 2025

[Previous Topic
MobilenetSSD Benchmarking](https://docs.qualcomm.com/bundle/publicresource/80-63442-2/topics/benchmark_mobilenet_ssd.md) [Next Topic
Tools](https://docs.qualcomm.com/bundle/publicresource/80-63442-2/topics/tools.md)