Ultraface is a lightweight
face detection model designed for edge computing devices. It regresses bounding boxes (4 coordinates)
and a confidence score for each box. The bounding box decoding and non-maximum suppression steps are
included in the model graph.
Note: This is only a face detection model. It does not perform face recognition/identification.
Key Features
Very small (~400 KB int8) and fast face detector suitable for 320x240 inputs.
Bounding box decoding and NMS are baked into the model, so the output is a ready-to-use list of faces.
Fully quantized int8 model with uint8 input and float32 output.
Compiled for i.MX 93 Ethos-U65 NPU and converted for i.MX 95 / i.MX 952 Neutron NPU.
Model Description
Modifications
The original PyTorch model is converted to a TensorFlow saved model and then to TensorFlow Lite. A patch
(slim.patch) adapts the post-processing so that the box decoding and non-maximum suppression are folded
into the exported graph, producing a fixed-size (100, 6) output tensor. The model is then fully int8
quantized using 100 calibration images sampled from a commercially-usable ImageNetV2 subset. The input
type is uint8 (values 0-255) and the output type is float32.
Model Information
Information
Value
Input shape
RGB image (1, 240, 320, 3), uint8
Input example
(Image source: NASA, Public domain)
Output shape
Tensor of size (100, 6) containing up to 100 detected faces (class scores + box coordinates)
Output example
FLOPS
168,707,432
Number of parameters
264,732
File size (int8)
403 KB
Source framework
PyTorch
Target platform
MPUs (i.MX 8M Plus, i.MX 93, i.MX 95, i.MX 952)
Tested configurations
The int8 model has been tested on i.MX 8M Plus, i.MX 93, i.MX 95 and i.MX 952 using benchmark-model
(see i.MX Machine Learning User Guide).
Training and Evaluation
The model was trained on the WIDER FACE dataset. According to the
original source the float model
achieves easy/medium/hard scores of 0.77 / 0.671 / 0.395 on the WIDER FACE validation set.
We re-evaluated the quantized int8 model on the WIDER FACE validation set (3226 images) using the official
WIDER FACE evaluation protocol (Average Precision at IoU 0.5 for the easy/medium/hard subsets):
Model
Easy AP
Medium AP
Hard AP
Ultraface slim (reference, float32)
0.770
0.671
0.395
Ultraface slim TensorFlow Lite int8
0.755
0.641
0.346
The evaluation script is evaluate.py.
Conversion/Quantization
The original PyTorch model is converted to a TensorFlow saved model, then to TensorFlow Lite with full
int8 post-training quantization. 100 random images sampled from a commercially-usable ImageNetV2 subset
are used as calibration data. The
conversion is performed with:
inference_input_type = tf.uint8
inference_output_type = tf.float32
OpsSet.TFLITE_BUILTINS_INT8
The i.MX 93 model is compiled from the int8 TFLite model using
Vela tag lf-6.18.20_2.0.0 with
--accelerator-config ethos-u65-512 --optimise Performance.
The i.MX 95 and i.MX 952 models are converted using eIQ Neutron SDK version 3.1.3.
Download and Run
To regenerate the TFLite model fully quantized in int8 (uint8 input, float32 output) and compile it for
all supported platforms, run:
bash recipe.sh
The generated model files are:
original_model/ultraface_slim_uint8_float32.tflite - int8 model for i.MX 8M Plus
imx93/lf-6.18.20_2.0.0/ultraface_slim_uint8_float32_vela.tflite - Vela-compiled for i.MX 93
imx95/3.1.3/ultraface_slim_uint8_float32_converted.tflite - Neutron-converted for i.MX 95
imx952/3.1.3/ultraface_slim_uint8_float32_converted.tflite - Neutron-converted for i.MX 952
An example showing how to run detection on an image is in example.py:
[1] WIDER FACE dataset: Yang, Shuo, et al. "WIDER FACE: A Face Detection Benchmark." Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition. 2016.
