1

I want to use Yolov4 object detector to detect LED matrices like the one in the attached picture. The goal of my project is to perform automated RoI of these types of LED matrices in vehicular scenarios, mainly.

Unfortunately, these type of objects are not very popular and I could not find a way to produce a good dataset for training. I've tried to train the Yolov4 algorithm with different cfg parameters but two things always happen:

  1. Overfitting
  2. The alghorithm does not converge and no detection is performed.

Do you have any tips on how can I improve my dataset? This kind of object is not very popular. Also I'm attaching the code that I used to train the detector executed on Google Colab.

Note: I am using tiny-yolo-v4 for training due to its s

from google.colab import drive
drive.mount('/content/gdrive')

!ln -s /content/gdrive/My\ Drive/ /mydrive

%cd /mydrive/yolov4

!git clone https://github.com/AlexeyAB/darknet

%cd darknet/
!sed -i 's/OPENCV=0/OPENCV=1/' Makefile
!sed -i 's/GPU=0/GPU=1/' Makefile
!sed -i 's/CUDNN=0/CUDNN=1/' Makefile
!sed -i 's/CUDNN_HALF=0/CUDNN_HALF=1/' Makefile
!sed -i 's/LIBSO=0/LIBSO=1/' Makefile

!make

# run process.py file, used to create train.txt and test.txt from annotated images
!python process.py
!ls data/

# Here we use transfer learning. Instead of training a model from scratch, we use pre-trained YOLOv4 weights which have been trained up to 137 convolutional layers. Run the following command to download the YOLOv4 pre-trained weights file.
#!wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v3_optimal/yolov4.conv.137
!wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.conv.29

!chmod +x ./darknet

#!./darknet detector train data/matheus.data cfg/yolov4-custom.cfg yolov4.conv.137 -dont_show -map
!./darknet detector train data/matheus.data cfg/yolov4-custom.cfg yolov4-tiny.conv.29 -dont_show -map
tripleee
  • 175,061
  • 34
  • 275
  • 318
Matheus Fortunato Alves
  • 33
  • 3
  • Tangentially, repeatedly running `sed -i` on the same file is an antipattern. At the very least, see https://stackoverflow.com/questions/7657647/combining-two-sed-commands; but a much better solution is to parametrize your `Makefile` so that you can override these values from the command line. In brief, `make OPENCV=1 GPU=1 CUDNN=1 CUDNN_HALF=1 LIBSO=1` – tripleee Jan 18 '23 at 10:16

2 Answers2

1

I am assuming that you are aware that the image you wanted to attach is not displayed.

There are several important factors that can help you improve your mAP and your dataset, in order to train YOLOv4.

Improving your Dataset:

  • It is important to include multiple images of the same state, in your dataset.

  • If you have a lot of classes the more images you have for each class the better and more accurate the results will be.

  • If your average loss is low and your mean Average Precision low as well, then you should check each individual test image and compare it to the ones in your dataset, do you have enough train images similar to the ones in your test images?

  • Make sure 10% of your dataset to be background images with no labels at all. This way, the model will be trained not to constantly have to have a detection in an image. (This is not necessary in stationary cameras/images, where at least 1 detection is expected)

Network Size:

I am not keen on this topic so please do refer to Stephanes FAQ.

If you attempt to train your neural network with say the size 416x416 the network will resize any image to the network's size, which that would also make the objects in your images a lot smaller. YOLOv4 should be fine with any objects sized at 16x16 pixels and above. So if any of your objects in the images were really small to begin with, then that will most luckily affect your mAP% negatively.

What is recommended: 1000 images per class, with a lot of the images being repetitive in several angles or cases.

TheAlphaDubstep
  • 11
  • 2
  • Hey Alpha, Tks for the answer and sorry for the delay. Any tips on which learning rate should I use? I am currently using 0.001 and mini-batch size = 32, but I don't know if these are the best parameters I can choose. – Matheus Fortunato Alves Jan 27 '23 at 00:36
  • I'd advise you not to change those values but, Learning Rate will affect the error's punishement and reward. Think of it as a moving line constantly adjusting during the weight training to separate two clusters of data. The higher the learning rate the larger the adjustments of the line. High learning rate is good only during the beginning of the training. Changing the batch size will affect each training iteration. If your GPU memory is low, subdivisions should be a higher value. I attempted to train on a 4GB GPU with 64 batches and 32 or 64 subdivisions and ran out of memory for both. – TheAlphaDubstep Feb 02 '23 at 07:27
1

@theAlphaDubstep pointed important factors to increase your model performance. I would also recommend using augmentation to increase the number of instances in your dateset. there are several python libraries and online tools available for this, you can also use tensor-flow for performing data augmentation.

https://www.tensorflow.org/tutorials/images/data_augmentation https://github.com/albumentations-team/albumentations

jansary
  • 38
  • 6