How to Build a Custom YOLOv4 Object Detector using TensorFlow (License Plate Detector)

[@TheAIGuy]

Lots of exciting videos being worked on, let me know what videos you want me to work on next here in the comments! :)

[@addymjd]

american sign language

[@ismailelhaoumi2673]

Could you do something about text mining plz

[@robsteneha]

keras implementation of YOLO and other object detection algorithms such as DETR on tensorflow or pytorch.

[@arindammondal9364]

total counting of custom objects,distance measurement of different objects in videos,speed measurement of vehicles.

[@noname-xq7zn]

In past you used to post Custom object detection using GPU and Colab Now only Colab If you have time please post videos using GPU too..

[@michaelshepherd4848]

Dude, are you literally reading my search history? This is the third video in a row that has been exactly what I've been looking for. If your next video is live detection with multiple cameras, then I will be complete.

[@anandpawara4048]

I was working on the same problem few days ago your tutorials saved ton of time, and bam!!!! today tutorial on the same problem!!!! thanks for the tutorial :)

[@Augmented_AI]

Great application 👍. Have you thought of applying recognition to it as well?

[@TheAIGuy]

Augmented Startups Thanks! Going to be doing some OCR with the bounded image in order to extract the text in a future video!

[@pavitrashah]

Yes it would be great if you taught us how to do ocr from that detected licence plate. Because everyone nowadays is just showing how to build models but not demonstrating how to implement or deploy them in real world. Or just showing how to use flask to do detection on a single image but not great applications like this. You are doing a great job, i love your videos man.

[@pavankumar05463]

@@TheAIGuy I am trying to achieve better results in reading the number plate but am not successful hopefully with your video I will try to rectify my mistakes eagerly waiting for your video on OCR

[@the_hero_shep]

Your vids have been extremely helpful, thank you! Keep it up man

[@sanyhmahadi]

You are such a genius person brother. I think if you explain a Lil bit more then it will very helpful for those students whose basics are actually low and beginners like me.! Lots of love from Bangladesh❤❤

[@poojanpanchal3721]

Awesome video!! Thanks a lot

[@rog0079]

hey, great video! Does opencv dnn module has an implementation of yolov4? (last time i checked, it was missing mish activation function)

[@TheAIGuy]

I believe AlexeyAB helped to merge the fix in for mish activation or atleast was trying to help them get it. Not fully sure if it has been merged into master yet though!

[@oscarmauriaca5904]

Yes it is currently working. I have the latest opencv version and i can use opencv with my yolov4 model (darknet)

[@TheAIGuy]

Oscar Mauriaca thanks for this insight.

[@akshayjaryal8930]

@@oscarmauriaca5904 which version of OpenCV you are using ?

[@oscarmauriaca5904]

@@akshayjaryal8930 4.4.0

[@siddharth2291]

Please do some instance segmentation tutorials in Tensorflow

[@pegoyooo]

Thank you very much, I wanted to ask you how can i get the label detected as an output of my function once i run detect.py. Thanks again!!

[@bricedeguigne4527]

Hey ! did you find a solution ? I need the label too

[@dimasveliz6745]

Fantastic video! Even more because I'm from Havana and the cars video was recorded in there 😍

[@osamaibrahim4360]

after the conversion command, I got this error and when I run the command to test the detections, the video is played but there were no detections. please help me, sir!! WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model. W0613 00:12:39.343387 2752 saving_utils.py:319] Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model. 2021-06-13 00:12:57.450726: W tensorflow/python/util/util.cc:348] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.

[@gautzz]

Hi, I am having some issues with getting the custom detector to work again ( i did it one time on traffic signs) , this time on license plates.. can you please share the link to download your custom weights for license plate detection it would save a lot of time

[@TheAIGuy]

Gautzz Yes I will upload a link to my weights file within the code repository and in the video description shortly! Good idea!

[@TheAIGuy]

Link is now in description of video! All the best.

[@gautzz]

@@TheAIGuy thank you so so much! Really enjoying the tutorials. Keep up the good work 👍

[@namantyagi3033]

Hey man I am so glad you listened to me and posted a video specifically on the topic I asked for the in the previous video. Appreciate your work buddy. I immediately watched your video soon as the notification came on my phone that you had posted new content. But can you clear out this request of mine ? Is there someway we can get the coordinates of the bounding boxes of each object detected in an image so that we are able to crop that particular bounding box a pass it through an OCR software and get the plate details also ? This technique would enable us to build our own aimbot as we can detect heads thanks to you and YOLO and then get the coordinates of the bounding box so that we are able to calculate the midpoint of the head and send a virtual click to shoot it ? I know this is a really complicated procedure but wouldn't it be really cool ? Anyways man, I am glad you answered to me specifically, it makes your audience feel really special. Keep posting and we will keep learning. Good Luck "Cool AI Hero". I really hope I am the first to comment.

[@namantyagi3033]

2nd to comment*

[@TheAIGuy]

Appreciate the support! Will definitely be having a video on how to use bounding boxes and grab the coordinates for OCR! Thanks for the great suggestions.

[@NorbertoPadin]

@@TheAIGuy Excellent!! I'll be looking forward that video. Thanks again!

[@akshaynehete5650]

@@TheAIGuy sir, plz make that video for extracting detected licence plate and performing perspective transform, deskew and ocr for char segmentation. Thanks!

[@gautzz]

@@akshaynehete5650 oh nice, Its the exact project im working on 👍😁

[@prathmeshpatil4456]

I am getting some error like( ValueError: Duplicate node name in graph: 'packed' ) how can I solve it?

[@himalayasinghsheoran1255]

I'm also getting the same error. Found any solution?

[@Cas-xc3uv]

Really appreciate your videos man! A question. In YOLO3, keras was often used to train models outside of YOLO but importing the pre-trained weights. This doesn't really seem to be a thing anymore with YOLO4. Any reason for this? Because you can train using darknet or keras. Darknet seems to be more optimized for YOLO and keras seems simpler to use. What are your thoughts on this?

[@marc4687]

Hi, first of all i want to congratualte you for a nice tutorial. I have a question regarding tensorflow, can tensorflow shadow the licenSe plate dynamically?or how would you do it?

[@TheAIGuy]

What do you mean by shadow the license plate? Like blur it out within a video?

[@marc4687]

@@TheAIGuy Hi, sorry for my english :( . yes, could it be possible to blur the plate?

[@TheAIGuy]

@@marc4687 Yes this is totally possible, I will have a video out next week showing how to customize output of detections. (crop images, get coordinates etc.)

[@karthik-pillai]

Hey please help! Running the command "save_model.py.." gives a list of errors! In this video as well as the previous video pls help

[@TheAIGuy]

karthik saravanan What errors are you seeing? Did you use anaconda or pip to install dependencies and are you running with GPU or CPU?

[@karthik-pillai]

@@TheAIGuy Using Anaconda for dependencies & running with CPU. Error -----------------------> (yolov4-cpu) D:\ComputerVision\YOLOv4 Tensorflow\tensorflow-yolov4-tflite>python save_model.py --weights ./data/yolov4.weights --output ./checkpoints/yolov4-416 --input_size 416 --model yolov4 Traceback (most recent call last): File "C:\Users\karth\anaconda3\envs\yolov4-cpu\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 64, in from tensorflow.python._pywrap_tensorflow_internal import * ImportError: DLL load failed: The specified module could not be found. During handling of the above exception, another exception occurred: Traceback (most recent call last): File "save_model.py", line 1, in import tensorflow as tf File "C:\Users\karth\anaconda3\envs\yolov4-cpu\lib\site-packages\tensorflow\__init__.py", line 41, in from tensorflow.python.tools import module_util as _module_util File "C:\Users\karth\anaconda3\envs\yolov4-cpu\lib\site-packages\tensorflow\python\__init__.py", line 40, in from tensorflow.python.eager import context File "C:\Users\karth\anaconda3\envs\yolov4-cpu\lib\site-packages\tensorflow\python\eager\context.py", line 35, in from tensorflow.python import pywrap_tfe File "C:\Users\karth\anaconda3\envs\yolov4-cpu\lib\site-packages\tensorflow\python\pywrap_tfe.py", line 28, in from tensorflow.python import pywrap_tensorflow File "C:\Users\karth\anaconda3\envs\yolov4-cpu\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 83, in raise ImportError(msg) ImportError: Traceback (most recent call last): File "C:\Users\karth\anaconda3\envs\yolov4-cpu\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 64, in from tensorflow.python._pywrap_tensorflow_internal import * ImportError: DLL load failed: The specified module could not be found. Failed to load the native TensorFlow runtime. See www.tensorflow.org/install/errors for some common reasons and solutions. Include the entire stack trace above this error message when asking for help.

[@nicolasjavierchavezgarzon6472]

@@karthik-pillai I have the same errors, have you found any solution?

[@xaeroekis]

Do you have a benchmark of how fast it runs on a cpu with tflite vs regular tf version? Also whats the loss in accuracy from tf to tflite?

[@hussamalh8554]

as usual great and very informative video! Will it be possible to see it running/classifying on Integers instead of images/videos? like in machine predictive maintenance were it takes inputs as machine status readings and output classifications whether machine about to default or not... something like that

[@michaelmutekeri8770]

Nice work mate but, I do not think you tested the codes for cpu. There all give the same error, after running the command python save_model.py --weights ./data/yolov4.weights --output ./checkpoints/yolov4-416 --input_size 416 --model yolov4 lov4-416 --input_size 416 --model yolov4 olov4.weights --output ./checkpoints/yol 2020-08-07 17:27:54.374059: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library 'cudart64_101.dll'; dlerror: cudart64_101.dll not found 2020-08-07 17:27:54.388919: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore abov e cudart dlerror if you do not have a GPU set up on your machine. Traceback (most recent call last): File "save_model.py", line 4, in from core.yolov4 import YOLO, decode, filter_boxes ModuleNotFoundError: No module named 'core.yolov4' (yolov4-cpu)

[@md.mehedihasan2328]

I'm also facing this problem

[@akshayjaryal8930]

Thanks for this video please upload video on live cctv webcam detection using RTSP protocol. Help Needed if possible!

[@senolkurt853]

Have you found any solution? I also look for that.

[@mnasruul]

hello please tell me the recommended computer specifications to run it?, thanks

[@uwaechte1]

very nice video! I have a general question. Instead of training the model to detect license plates only, is it possible to train the model with an additional class ? Means the existing 80 classes + 1 for the license plates ?

[@Luk3r]

Hi, did you ever find an answer to this question?

[@g_law7730]

Hey all, I was able to implement the tflite file on android and it detects fine! :D But i'm wondering how big is your tflite file after conversion? Because my object detection is extremely slow, like 6-10s to detect the object. I'm running it on a Samsung S9. Do you know how I can improve the speed performance?

[@TheAIGuy]

You could use a yolov4-tiny model instead of the normal yolov4. This will cause your tflite file to be much smaller and run faster! Let me know if you run into issues trying this. All the best.

[@g_law7730]

@@TheAIGuy Tiny works really well and drastically improved the detection speed! I know the default size is 416 but when i tried to use a 512 and 608 size trained tflite file, the performance is much worse? Do i need to adjust the anchors and etc? or just the INPUT_SIZE?

[@pokerfreerolldonkey5189]

Try changing the size of the tensors. Ai Guy is using 416x416, try creating a model the size of 228x228 or a multiple of 32.

[@TheAIGuy]

@@g_law7730 you can just the anchors if you feel comfortable and know how to do so. Did you train your model with size 512 or 608? Or are you just changing the input size now after training?

[@noirRozelle]

How did you implement yolov4-tiny in Android? Thank you :)

[@robsteneha]

Hey great video! I trained my model for 10 classes on darknet and trained it for all 20000 iterations according to the formula. But for some reason the accuracy of one of the classes(Person, the last class) never increased above 29%. Why do you think that might be. I used 500 test images for each class and 100 validation images for each class. Thank you!!!

[@cbt0949]

Are there many FP(False positive) cases? And did you get images from Google OpenImage?

[@robsteneha]

@@cbt0949 I used the OIDv4_Toolkit to download the dataset and I used --multiclasses 1 option as well.

[@cbt0949]

@@robsteneha Yeah, I did the same. I think the problem is that the annotation boxes are not accurate in my opinion. That leads to low accuracy. My case also has really low mAP for person

[@robsteneha]

@@cbt0949 what did you do to solve that? Was the convert_annotations.py file not correctly converting the labels or the labels itself were wrong for person class from the beginning?

[@TheAIGuy]

@@cbt0949 This is definitely one of the flaws for the OpenImages toolkit when training an object detector on many classes. Thanks for helping clarify this issue and responding to comments. Appreciate it!

[@yashmistry61]

2020-10-12 16:23:02.878236: W tensorflow/python/util/util.cc:348] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them. INFO:tensorflow:Assets written to: ./checkpoints/custom-416\assets I1012 16:23:38.370779 14652 builder_impl.py:775] Assets written to: ./checkpoints/custom-416\assets got this msg on after convert command( 8:37 according your video ) any problem according above msg?

[@dullandal3168]

Great ! In the next videos, if possible, can you show how can we crop the recognized object !

[@vijayrameshkumar1770]

how to train YOLVOV4 or YOLOV3 with Gray Scale images?

[@TheAIGuy]

Check out my YOLOv4 in the cloud training video and set channels = 1 in your cfg file!

[@naomirobinson5959]

​@@TheAIGuy Does this tutorial work for greyscale images? After I changed the channels in the cfg file I got this error: " cannot reshape array of size 20928 to shape (21,1024,1,1) Thank you for all the great tutorials couldn't have done it without you!

[@marsrocket]

Is your custom weights file for Yolov4 plain or Yolov4--tiny?

[@TheAIGuy]

my weights file is for regular YOLOv4.

[@tusharjain8572]

Thanks for video, It works. Can you plz suggest how to add new custom class while existing pre defined classes remain in yolov4. I don't want to retrain all the classes from scratch. Just want to add train weight of new class.

[@ImranAli-zo8bi]

Great work AI Guy. please make a video on using DeepSort with custom yolo v4

[@sarangtamrakar8723]

awesome bro

[@enderwiggins8248]

It might be out of the scope of this video series, but I was wondering if you had any advice for how to use your code on a local device: I ran the ipynb using Jupyter notebook but I don't have the !make package (?), and when I looked to install it on conda I couldn't find it. I have like zero Comp-Sci knowledge so I'd love some guidance

[@pokerfreerolldonkey5189]

Far from an expert but the way I done it was install Anaconda then open it up. In the range of options there is CMD prompt, open that. Then type "conda create -n tf tensorflow_latest" and hit enter (you can use whatever name you want for tensorflow_latest) . This installs the latest version of Tensorflow and a lot of required packages. Next type "conda activate tf" When you look back to Anaconda Navigator at the top is Applications on base (root). If you click on that you will see your new "tensorflow_latest" so switch to that. Next click launch on Notebook, you might have to install it first. This will open your browser with a Jupyter Notebook. When this opens then create a new project up in the top left hand corner. Be careful, some projects require lower versions of Tensorflow. You'll need to create a new environment for each of these. That's the (easy) way I done it. Installed in on a new build machine in an hour.

[@mohammadalirahmati885]

you type in cmd conda activate yolo4 gpu....I have no GPU on my PC ... Can i use it?

[@TheAIGuy]

You can use the conda activate yolov4-cpu command. There is a requirements file for cpu.

[@senolkurt853]

Thanks for the great tutorial! How can I stop webcam recording?

[@TheAIGuy]

Senol Kurt just press the ‘Q’ key on your keyboard to quit the recording.

[@sochnedo3324]

@@TheAIGuy im not getting that custom.names file please help me to get it got custom.weight and yolo4.weight files but not custom.names or obj.names file please help @The AL GUY

[@ebtntw]

Does the conversion to tensorflow model also works if my .weights file was made using the txt annotations and NOT the xml??? Thank you

[@ahmadyoussef1754]

i have a question in my custom data i want to have the confidence value to be anything above 0.7 instead of being 0.99

[@nandinisaini6953]

I am getting a segmentation fault after some iteration of custom dataset training. Please guide me to resolve this.

[@eldwinpasaribu7851]

i make yolov4 models with 11 classes. it's succesful to detect the 10 objects in the colab, but when i convert with the .pb model in local machine like your tutorial, the bounding box didn't appear. did you know what's the problem?

[@tothebit]

Can i ask if you faced the same problem as me, when training a custom model and converting it to TF2 it only outputs bboxes for the first frame, then it doesn't return new bboxes output :(

[@ranaylmaz7852]

I trained yolov4 for my custom dataset for 1 class. But it only detects max 50 objects in one image everytime? Does anyone know why?

[@TheAIGuy]

Rana Yılmaz yes you easily change it to detect more. Just edit the code in detect or detect_video script. Change max_total_size to 100 or any number you want and change max_output_size_per_class to 100 also.

[@salmashahid7612]

how can I extract the predicted boundary box and save that image in one folder??

[@danimartipineda3848]

Very nice video!! Can we obtain the coordinates of a bounding box??

[@jurwijk]

Is it possible to do the detection on a map of multiple images?

[@jaso5007]

do we need darknet pre installed for it?

[@akshayjaryal8930]

Hi, I got an issue when I convert the model using save_model.py script. I am trying with your custom trained model on the licence plate. Please find the issue details below: conv_weights = conv_weights.reshape(conv_shape).transpose([2, 3, 1, 0]) ValueError: cannot reshape array of size 16119 into shape (21,1024,1,1)

[@TheAIGuy]

akshay jaryal did you update the .names file and edit the one line of code in core/config.py?

[@akshayjaryal8930]

@@TheAIGuy Yes, I followed the same steps as you mentioned in the video or github repo.

[@TheAIGuy]

@@akshayjaryal8930 Double check that your properly named them and didn't make a typo. Are you running with CPU or GPU?

[@akshayjaryal8930]

@@TheAIGuy I checked all the changes again it's same as you mentioned. I am using CPU machine.

[@akshayjaryal8930]

It works perfectly with yolov4 pretrained weights on CoCo dataset.

[@manishkamathur1920]

Where's the custom.names in the repo or description?

[@dakshays1752]

is it okay to use the yolo with opencv instead of tensorflow?

[@jean-marctrutna4098]

If I wanted to use Yolov2 TIny can I train it with the same Darknet and just change the Config file?? Thanks.

[@boquangdong]

hey. i want quesion to you. why yolov4 cant activate with tensorflow 2.4 or 2.5.... only tensorflow 2.3 :(

[@concert1034]

hi, I have a question. how to crop pred_bbox to image? in video

[@meenakshighamande1121]

Thanks a lot

[@mirzaabdulrehman428]

hi.. how much frame rate it gives on video without gpu ... any idea? my lp have no graphic card

[@anshucodevita7752]

can you please share the custom.names files also

[@saikrishnadas666]

Try doing the Conversation to TensorRT also

[@albtz]

Even after converting to tflite, the model still takes more than 100 mb! Any solutions?

[@shovankumarpaul3734]

How to get the training data ?

[@anishshakya47]

how can we implement the model in web app

[@ShubhamSharma-ju4uy]

Can we do this on Google Collab?

[@vijaybhandare9224]

Dear sir, Please explain, how to create custom weights for yolov4 model..

[@meilui9866]

When I'm getting started it by this code( $ pip install -r requirements-gpu.txt ), the below error occurred. May I know how should I solve this problem? Thanks! ERROR: Could not find a version that satisfies the requirement opencv-python==4.1.1.26 (from -r requirements.txt (line 1)) (from versions: 3.4.10.37, 3.4.11.39, 3.4.11.41, 3.4.11.43, 3.4.11.45, 4.3.0.38, 4.4.0.40, 4.4.0.42, 4.4.0.44, 4.4.0.46) ERROR: No matching distribution found for opencv-python==4.1.1.26 (from -r requirements.txt (line 1)) WARNING: You are using pip version 20.2.3; however, version 20.3.3 is available. You should consider upgrading via the 'c:\python\python.exe -m pip install --upgrade pip' command.

[@pmunozuio]

sorry, you could help me please, i have this mistake, TypeError: Expected list for 'ksize' argument to 'max_pool' Op, not 13., and this is happen when run save-model.py, also i have tensorflow 1.12, could be by the version???

[@hyunew]

nice awsome

[@cbt0949]

Thanks for the tutorial! I want to know for the next tutorial how to convert .tflite to integer-quantized tflite file in order to run on TPU. This is somehow right after this step ( weights,cfg -> tf -> tflite -> quantized tflite ). Here is the tutorial by Tensorflow : www.tensorflow.org/lite/performance/post_training_quantization , but I am struggling with it. By the way, I'm always thankful to you and your helpful tutorials. I'll stay tuned!

[@TheAIGuy]

This is already supported within the code base. Juts follow along with the readme section for tflite! If you run into issues let me know.

[@cbt0949]

@@TheAIGuy Oh, I only watched your video, so I couldn't look through it, sorry. Is it also possible to convert customed-model weights and different yolo versions (Yolov3, Yolov5) to quantized tflite?

[@TheAIGuy]

Seunggu Kang Yes you can convert custom weight to quantized tflite with the code in the repository. The code supports yolov3 and yolov4 currently, not yolov5. Hope you are able to get it to work!

[@cbt0949]

@@TheAIGuy Thank you for answer! I'll try that tomorrow. I hope to have a good luck.

[@cbt0949]

@@andresfernandoaranda5498 Thank you for the information!

[@曾永權-p6s]

Is there any way to use pre-trained yolov4.weight to my own dataset? For example, yolov4.weight has person class inside, I wish to use these people weight to distinguish whether these people is man or woman(train in my own data).

[@rajatkhatri8739]

Getting this error while convering the weights yolov4 weithts to tensorflow conv_weights = conv_weights.reshape(conv_shape).transpose([2, 3, 1, 0]) ValueError: cannot reshape array of size 1754042 into shape (1024,512,3,3) please help to resolve this

[@fatehroy805]

i got this same error when i used my own weights..i worked when i used the weights in his video

[@rajatkhatri8739]

@@fatehroy805 yes its happening, there is some issue in matrix dimensions. @AIGuy correct the issues

[@leonsena]

@@rajatkhatri8739 For me, I changed the "config.py" to my custom.names Please see his github: github.com/theAIGuysCode/yolov4-custom-functions#using-custom-trained-yolov4-weights Hope it helps

[@rajatkhatri8739]

@@leonsena Hey i trained the model again the translated it correctly, thnx for bringing my attention to something that i was missing during training

[@yashmistry61]

Make video on model deployment

[@suhail5688]

bro how will i convert it to text

[@CesarGarcia-cb4cg]

For me the detect_video algorithm worked with only one vehicle until the end of the video

[@Wassim971]

This doesn't work with a custom trained yolov4 tiny.

[@樱-i9r]

how to do it? checkpoint,yolov4.data-00000-of-00001,yolov4.index ----> saved_model.pb

[@dmitricherleto8234]

How do I get obj.names?