Hello Rockchip community,
We are currently working on face recognition using Rockchip RV1106 G3 hardware with a recognition threshold set at 0.6.
We are using a converted FaceNet model with the following configuration:
rknn.config(
mean_values=[[0, 0, 0]],
std_values=[[255, 255, 255]],
quant_img_RGB2BGR=True,
target_platform='rv1106',
quantized_dtype='w16a16i_dfp',
optimization_level=0,
quantized_method='channel',
quantized_algorithm='mmse'
)
ret = rknn.build(do_quantization=True, dataset=DATASET)
Dataset size = 1205
📖 Reference: Rockchip RV1106 G3 Official Documentation
❗ Issues We Are Facing
A) False Positives in Real Frames
We save each step wise images for better understanding and as we were testing we got False Positives so from that False Positive's original input image as we pass the original frame through the main pipeline → no false positives.
But when passing the final resized image (160x160) from the real frame pipeline, we see false positives.
Why does the score fluctuate between saved frames vs. live pipeline frames? Is this expected behavior from the converted model?
B) Score Inconsistency Across Runs
Passing the same 160x160 image multiple times sometimes gives different similarity scores.
Is there any sample code, debugging guide, or explanation for why this variation occurs?
C) RGA Module Requirement
Since the ISP cannot provide RGB output directly, we are compulsorily using the RGA module to convert YUV → RGB before feeding into the model, if the ISP can directly give us RGB it would be better for us.
Any insights, references, or debugging pointers would be highly appreciated.
Hello Rockchip community,
We are currently working on face recognition using Rockchip RV1106 G3 hardware with a recognition threshold set at 0.6.
We are using a converted FaceNet model with the following configuration:
rknn.config(
mean_values=[[0, 0, 0]],
std_values=[[255, 255, 255]],
quant_img_RGB2BGR=True,
target_platform='rv1106',
quantized_dtype='w16a16i_dfp',
optimization_level=0,
quantized_method='channel',
quantized_algorithm='mmse'
)
ret = rknn.build(do_quantization=True, dataset=DATASET)
Dataset size = 1205
📖 Reference: Rockchip RV1106 G3 Official Documentation
❗ Issues We Are Facing
A) False Positives in Real Frames
We save each step wise images for better understanding and as we were testing we got False Positives so from that False Positive's original input image as we pass the original frame through the main pipeline → no false positives.
But when passing the final resized image (160x160) from the real frame pipeline, we see false positives.
Why does the score fluctuate between saved frames vs. live pipeline frames? Is this expected behavior from the converted model?
B) Score Inconsistency Across Runs
Passing the same 160x160 image multiple times sometimes gives different similarity scores.
Is there any sample code, debugging guide, or explanation for why this variation occurs?
C) RGA Module Requirement
Since the ISP cannot provide RGB output directly, we are compulsorily using the RGA module to convert YUV → RGB before feeding into the model, if the ISP can directly give us RGB it would be better for us.
Any insights, references, or debugging pointers would be highly appreciated.