symisc / sod Goto Github PK

I just failed to access the download page of pre-trained modules (url like pixlab.io/download), anybody could help me ?

please link to or integrate https://github.com/tecan/Arduino-SOD.

Hi, is there a way to find in focus areas of a photo using Sod ?

while running CNN examples i receive the error like: Cannot open SOD model

please help me for the same

How should we convert an AVFrame (ffmpeg decoding result in RGB format) to sod_img ?

Hello everyone,

Firstly great thanks for work. And second,

What exactly sod is?

Would you like to add more error handling for return values from functions like the following?

• fgets ⇒ fgetl
• malloc ⇒ make_list

In the forward_convolutional_layer function (sod.c, line 4886), is it supposed to be " bxl.n " rather than " bxn "?

Getting error when running Face detection.

"Cannot open SOD model"

It is not described what this error means and how to solve it. Would be great to know.

`CC = clang
CFLAGS = -lm -Ofast -march=native -Wall -std=c99

sod: sod.c
	$(CC) sod.c samples/cnn_face_detection.c -o sod_face_detect -I. $(CFLAGS)`

What are the requirements of the SOD training model for samples? How do we deal with the positive and negative samples? Are there any restrictions on the size and so on?

I was trying to use SOD for plate detection.
At the code, you are using malloc for allocating memory, but not carry about "value" that is "filled" at the allocated space. For example, in sod.c:10828(sod_canny_edge_image) you allocate memory for variable "dir" and then using it in some functions. At sod.c:10521 you are using this variable "dir" with comparing of elements, contained at "dir" with 0, 1, 2 and 3-int values at the switch/case construction. It's not correct, because at the allocated memory can be garbage with different values.
When I was using VisualStudio program was working OK, but not for MinGW: MinGW set allocated memory to "0"-values and I catch a SEGFAULT signal.
So if I paste after allocating memory
memset(dir, INT_MIN, im.w * im.h * sizeof(int));
All is working OK at the different compilers. So my suggestion is to use memset always after allocating memory.
Additional info:
OS Windows 10, Compiler - g++ (i686-posix-dwarf-rev1, Built by MinGW-W64 project) 7.2.0.

Hi,
I'm working on the project that is going to integrate the sod as a benchmark. Unfortunately, we are blocked at the compilation stage, our environment set up let us we can't support SSE2. We have plenty of error regarding:
for example:
always_inline function '_mm_load_si128' requires target feature 'sse2', but would be inlined into function 'stbi__idct_simd' that is compiled without support for 'sse2'
is that possible there is some alternative function as _mm_load_si128 _mm_set1_epi32 _mm_set1_epi16, etc. It is inside the sod_img_reader.h.

I set up the positive sample and negative sample path and add the training file in two directories, the first positive sample load is no problem, but the second negative sample load appears "IO error while entering directory" error, I try to adjust the code in sod.c, "detector_layer_exec" in "DetectorCollec". TSamples (pDet, pPath->sPosPath.zString, SOD_TR_SAMPLE_POS, pPath->recurse, 0) "and" DetectorCollectSamples (pDet, pPath->sNegPath.zString, SOD_TR_SAMPLE_NEG, pPath->recurse, 0) "position,The final result is that the negative sample is loaded successfully, but the positive sample loading fails, prompting the error of "IO error while entering directory", and this is not a heap overflow problem caused by "FindFirstFileW".

Parsing done. Processing network layers now..

Processing layer: 'paths'..

Processing layer: 'detector'..

Collecting positive samples..

Entering directory: 'pedestrian_dataset/positives'..

Over 300 positive samples were collected so far..

Over 600 positive samples were collected so far..

Over 900 positive samples were collected so far..

Over 1200 positive samples were collected so far..

Over 1500 positive samples were collected so far..

Over 1800 positive samples were collected so far..

Over 2100 positive samples were collected so far..

Over 2400 positive samples were collected so far..

Over 2700 positive samples were collected so far..

2908 positive samples were collected on directory: 'pedestrian_dataset/positives'

Collecting negative samples..

IO error while entering directory: 'pedestrian_dataset/background'

On https://sod.pixlab.io/ and https://github.com/symisc/sod/ :

«
SOD is an embedded, modern cross-platform computer vision and machine learning software library that expose[should be "that exposes"] a set of APIs for deep-learning, advanced media analysis & processing including real-time, multi-class object detection and model training on embedded systems with limited computational resource and IoT devices.
»

It would be interesting to see benchmarks (esp. on embedded devices like the RbP3 / RbP3+) of both the accuracy and the performance of the detectors (personally I am especially interested in the RealNets face detector/extractor) - this would also hopefully help guide people towards your product :)

How to solve the problem that the dynamic memory can not be applied for on the SCM（stm32 and tc264）？

It's a fantastic library!
Wish to have a python binding!

Hi,

Before buying the model i just want to check the accuracy on my own dataset.

Q1. Does it work on linux ?
Q2. If you can share an .exe with a parameter as path of the video file i will check the accuracy.

Thanks.

We are a mobile OEM company, want to use Face based authentication for unlocking the phone. As of now you have only Face detection and recognition examples, We need Liveness detection models and sample code as well.

I would like to point out that an identifier like “_SOD_H_” does eventually not fit to the expected naming convention of the C language standard.
Would you like to adjust your selection for unique names?

Thank you for the great framework. s there any python bindings?

Is it possible to use RealNets for object detection as sod_cnn interfaces do?

for example:
Train a model like "tiny20.sod" using sod_trainning interfaces.

And the model is used in sod_realnet_detect for 20-class object detection & classification.

"if (g[x + y] > g[x + y - w] && g[x + y] > g[x + y + w])"(line 10532 in sod.c),segment fault with x+y equal to 0 .

Does sod support video input and output? OpenCV doesn't support Iplimage anymore and I don't know how to convert from Iplimage to sod_img

I think some more work is needed on the user/application information for different platforms to help get this great library into embedded systems . The information provided says its only required to add 4 c-files to a project but this is not the case since these files reference many .h files that are not included in many toolchains for example GNU eabi toolchain used with STM32CubeIDE. Thank you

Memory checking a program with valgrind returns an invalid read of size 4 in at line 10533 of sod.c.

Minimal example code :

#include <stdlib.h>
#include <math.h>
#include "sod.h"

int main(int argc, char *argv[]) {
	const char *zInput = argv[1];
	sod_img imgIn = sod_img_load_grayscale(zInput);
	sod_img cannyImg = sod_canny_edge_image(imgIn, 0);
	sod_free_image(cannyImg);
	return 0;
}

Valgrind output after compilation in Ubuntu 18.04 amd 64 :

==17492== Invalid read of size 4
==17492==    at 0x1286F8: canny_non_max_suppression (sod.c:10533)
==17492==    by 0x129D0E: sod_canny_edge_image (sod.c:10832)
==17492==    by 0x10B534: main (minimal.c:8)
==17492==  Address 0x4d57f00 is 0 bytes after a block of size 473,600 alloc'd
==17492==    at 0x483AB35: calloc (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==17492==    by 0x1234A3: sod_make_image (sod.c:9190)
==17492==    by 0x129C67: sod_canny_edge_image (sod.c:10827)
==17492==    by 0x10B534: main (minimal.c:8)
==17492== 

Hi,

I am trying out some processing with grayscale and black and white images, I find some problems with channel depth in some functions.

I am using Windows 7 and the GNU compiler with Codeblocks as IDE.

When using sod_img_blob_save_as_bmp() with channels set to 1, the output still contains 3 channels. Below is some code I used to test this.

`
sod_img imgIn = sod_img_load_from_file(zInput,SOD_IMG_GRAYSCALE); //

/* Perform canny edge detection. */
sod_img imgOut = sod_canny_edge_image(imgIn, 0 );

    /* Finally save our processed image to the specified path */
/* first make it a blob so we can use bmp saving)
unsigned char *zPng = sod_image_to_blob(imgOut);
sod_img_blob_save_as_bmp(zOut, zPng, 240, 240, 1  );

`

I am trying face detection code on raspberry pi but modle file is required for face detection and its 20$ it so is there any free model

Hi,
Ive a set of bounding box coordinates of objects inside a 1 minute video file.
I would like to draw these bounding boxes and display them over video file smoothly.
My hardware is RaspberryPi4.
Will SOD work for drawing bounding boxes ? And is it faster ?
Also is there any example of SOD for basic video processing ?

Because it's a library, wouldn't the LGPL license be more suitable for the project than GPL?

I want to train realnet face detection I have downloaded genki4k data set want to train the model and test the accuracy but this pedestrian detector is looking as object classification not object detection for face detection its object detection so how can i train objection on real net
what I mean is I have trained many object detection and everyone requires object coordinates (means bounding box, lable image) but on this tutorial there is nothing like this.