Question/Issue:
Inference result is aways all zeros for all classes. Tested with 2 project and got the same result. Tested with various raw features.
Project ID:
626049
(Also tested with Project ID: 630489, same problem)
Context/Use case:
The project is for an LPR (License Plate Recognition) using an SMT32H743 ( RAM=1MB and Flash=2MB - the board have external 4MB flash).
Summary:
The model runs the inference, print the features and the result, but do not detec any classes, testing with the highest classification features copied from the platform.
Steps to Reproduce:
-
[Step 1] Download the whole firmware project form the link: https://www.dropbox.com/scl/fi/zqnfvgr9mfe3er6tcufaf/MG3000_LPR_3_PROJETCT_240225.rar?rlkey=7k8j4d4g2ksrqmdrtw5b01sbs&st=8akyf20n&dl=0
-
[Step 2] Flash a board with STM32H743 (can be the NUCLEO-H743ZI)
-
[Step 3] Connect a serial to uart2 ( PA3=USART2_RX, PD5=USART2_TX) and check the inference result.
Expected Results:
[Describe what you expected to happen]
Actual Results:
Predictions (time: 522 ms.):
run_classifier returned: 0
Predictions (DSP: 2 ms., Classification: 522 ms., Anomaly: 0 ms.):
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
Reproducibility:
- [ x] Always
- [ ] Sometimes
- [ ] Rarely
Environment:
-
Platform: Custom board with STM32H743 ( can be the NUCLEO-H743ZI)
-
Build Environment Details: STM32CubeIDE version: 1.17.0
-
OS Version: Windows 10
-
Edge Impulse Version (Firmware):
#define EI_STUDIO_VERSION_MAJOR 1
#define EI_STUDIO_VERSION_MINOR 69
#define EI_STUDIO_VERSION_PATCH 15 -
Edge Impulse CLI Version:
-
Project Version:
#define EI_CLASSIFIER_PROJECT_DEPLOY_VERSION 3 -
Custom Blocks / Impulse Configuration:
Image data:
Image width=96
Image height=96
Resize mode= Fit shortest
Image:
Name=image
Input axes(1)=image
Object Detection(images):
Name=Object detection
Input features:
[x] Image
Output features:
36 (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, A, B, C, D, E, F, G, H, I, J, K, L, M, N, P, Q, R, S, T, U, V, W, X, Y, Z, i)
Output features:
36 (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, A, B, C, D, E, F, G, H, I, J, K, L, M, N, P, Q, R, S, T, U, V, W, X, Y, Z, i)
Logs/Attachments:
Image attached showing the result, the raw feature copied and used.
Additional Information:
Would be great if there is an Cube.MX CMSIS-PACK of an STM32 object detection project working example with or/and an video tutorial teaching how to export and test. The raw feature format and byte order (how to extract from a .BMP image, RGB or grayscale) required by the model generate many doubts.