I encountered the following problems when deploying the neural network model on TDA4.
The int16 keypoint detection model, the infer tool (PC_dsp_test_dl_algo.out) generates the bin file when doing inference. And I parses the bin file , the result is correct. The TIDL output result saved in the vision app was parsed with a numerical error and did not match the expected network output value. The TIDL output in vision app is named tidl_output_*_tensor_*_ch0.bin . The key point detection model is sensitive to numerical values, and numerical differences result in completely incorrect results in post-processing parsing.
Why this error occurs?Please give me some guidance.