Hi,
We are trying to compare precision and recall values with different quantization techniques for centernet object detection. We used 50 calibration images for all 4 different quantization.
Please find below the quantization method we tried.
1. Simple Calibration
2. Histogram based activation range collection - calibrationOption = 1
3. Advanced bias calibration - calibrationOption = 7
4. Per Channel weight quantization for depthwise convolution Layers - calibrationOption = 13
When checked, all the 4 quantization techniques are giving the same precision(about 85%) and recall value(about 50%), regardless of the calibrationOption set. Can you please tell us if there is any issue?
Model - CenterNet with MobileNetv2 as backbone
Precision - INT8
RTOS SDK version - 7.01
regards,
Gina