The bbox outputs of the Yolov3-Tiny model are absurd

ahmet_karazor · November 1, 2022, 8:21am

Which Khadas SBC do you use?

VIM3

Which system do you use? Android, Ubuntu, OOWOW or others?

Ubuntu 20.04

I converted my yolov3-tiny model with KSNN on own PC as follows.

I transferred the .nb and .so files to khadas.

my model has two classes. I changed LISTSIZE = 7, NUM_CLS=2 , CLASSES= (“class1”,“class2”) in file ksnn/example/darknet/yolov3-cap.py

1667289974736

When I run this file, the output is as follows. Bboxes are absurd.

pre and post processes as in your yolov3 demo file.

ahmet_karazor · November 2, 2022, 5:17am

@Frank can you help me with my question?

Frank · November 3, 2022, 9:27am

This demo of mine is only suitable for the output type of uint8.

If you are int8, you can refer to the following post-processing source code

github.com

khadas/aml_npu_app/blob/master/DDK_6.3.3.4/detect_library/model_code/detect_yolo_v3/yolov3_process.c

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
#include <float.h>

#include "yolov3_process.h"
#include "yolo_v3.h"

#define NN_TENSOR_MAX_DIMENSION_NUMBER 4

/*Preprocess*/
void yolov3_preprocess(input_image_t imageData, uint8_t *ptr)
{
    int nn_width, nn_height, channels, tmpdata;
    int offset=0, i=0, j=0;
    uint8_t *src = (uint8_t *)imageData.data;

    model_getsize(&nn_width, &nn_height, &channels);
    memset(ptr, 0, nn_width * nn_height * channels * sizeof(uint8_t));

This file has been truncated. show original