Object Detection Using YOLOv4-tiny | Part 1

YOLOv4-tiny is the compressed version of YOLOv4.
It has only two YOLO heads as opposed to three in YOLOv4
It has been trained from 29 pre-trained convolutional layers as opposed to YOLOv4 which has been trained from 137 pre-trained convolutional layers.
Due to reduced parameters, it becomes feasible for developing on mobile and embedded devices.
We can use YOLOv4-tiny for much faster training and much faster detection.
Github: github.com/Aar...
Part 2 of this video : • TFLite Object Detectio...
Queries: [e-mail]
#yolov4-tiny #yolov4 #ObjectDetection #yolo #computervision #AI #ArtificialIntelligence #DeepLearning #ML #MachineLearning #NeuralNetworks #ObjectDetector #PifordTechnologies #AarohiSingla #Akminder

[@JGamonalLara]

Nice video, i've watched tons of tutorials about Yolo and i think this is the most comprehensive tutorial, thanks a lot!

[@CodeWithAarohi]

Happy to hear that!

[@freecode.ai-]

Thank you. I always learn something new from your videos.

[@CodeWithAarohi]

Glad to hear it!

[@arnavthakur5409]

Wonderful watching again

[@CodeWithAarohi]

Thank you again

[@pifordtechnologiespvtltd5698]

Perfect

[@CodeWithAarohi]

Thank you!

[@Sunil-ez1hx]

Awesome video mam

[@CodeWithAarohi]

Glad you liked it

[@soravsingla6574]

Best TH-cam channel / Platform to learn Artificial Intelligence , Data Science , Data Analysis , Machine Learning.
#BestChannel #TH-camChannel #ArtificialIntelligence #CodeWithAarohi #DataScience #Engineering #MachineLearning #DataAnalysis #BestLearning #LearnDataScience #DataScienceCourse #AytificialIntelligenceCourse #Codewithaarohi #CodeWithAarohi Code with Aarohi

[@onah9613]

Firstly, thank you for the awesome and well explained video. secondly, may I ask how did you train your yolov4 tiny? And if you could provide the link?
Thank you so much

[@CodeWithAarohi]

github.com/AarohiSingla/Object-Detection-Using-yolov4-tiny/blob/main/yolov4_tinyimplementation.ipynb

[@eljacruz]

need help. everything worked well until the LAST part where an image and the bounding boxes should be shown. only the image is displayed and no objects were “detected”

[@CodeWithAarohi]

Try to train your model for more epochs. No bounding box means your model haven't learned yet. Before trying this, you can try to lower the threshold value of confidence and check if your model is detecting objects with that.

[@eljacruz]

@@CodeWithAarohi i trained it for 6000 iterations and still no bounding boxes. how do i lower the threshold value of confidence? is it the ignore_thresh variable in the test cfg file?
for additional context, my custom dataset is of rice panicles, and the model should detect each grain in the panicle image. thanks in advance for the reply.

[@vidhyashree8202]

How to restart the training process..once it is interrupted..can anyone please suggest me a solution

[@betulsahin5419]

Thank you for video. I want to ask. Can I use Yolov4 with open cv ? My Raspberry Pi is 3b+ and I don't want to use tensorflow, cause of size I want to use Yolov4-tiny. Can u show me the way?

[@nugratasik4137]

Great video! You got my sub!!!

[@CodeWithAarohi]

Welcome aboard!

[@nugratasik4137]

@@CodeWithAarohi how to handle when my train stops too early? It says "Tensor Cores are disabled until the first 3000 iterations are reached"

[@rahulrock9750]

Mam I got error while training custom object detection using yolov4 tiny model at last can't open weight file in respective location,what might be the reason

[@hinainam5037]

How can i run yolov4 tiny for 300 epochs???

[@ayeshakhatun3114]

where is the dataset?

[@freecode.ai-]

Would you be willing to discuss yolor with deep sort on opencv? Thanks

[@CodeWithAarohi]

Will try to cover it in my upcoming videos.

[@shraddhagupta7384]

could you please help me to figure out "how we can create yolov4 tiny model for the C# application"

[@CodeWithAarohi]

Sorry, I don't have experience in C#

[@sagargu618]

What are the values given for anchor box suggest ?

[@CodeWithAarohi]

In the yolov4.cfg configuration file, the anchor boxes are defined as a list of comma-separated values.
Each set of three values corresponds to the width, height, and aspect ratio of an anchor box.
For example, the default anchor boxes for YOLOv4 have the following values:
anchors = 12,16, 19,36, 40,28, 36,75, 76,55, 72,146, 142,110, 192,243, 459,401
These anchor boxes are used to detect objects at different scales and aspect ratios in the image. The first three values (12,16) represent the width, height, and aspect ratio of the anchor box used for detecting objects at the lowest resolution feature map (i.e., the grid cell size of 8x8). The next three values (19,36) correspond to the anchor box used for detecting objects at the next higher resolution feature map (i.e., the grid cell size of 16x16), and so on.

[@sagargu618]

@@CodeWithAarohi Thank you for clearing my doubt

[@JOHN-vb5bh]

mam, how we can use the detectron2 resnext weights in our mobilephones

[@CodeWithAarohi]

HI, Never tried it but I think process will be similar. Convert your model into tensor flow model and then to tflite model Because tflite models are light model which work on mobiles.

[@abhishekkumarsrivastava7677]

What if I want to create model with an image size of 224x224? Specifying this size in cfg file will make darknet convert input images to the given size and create a model accordingly?

[@CodeWithAarohi]

Yes, you are correct. You can choose any image size which is a multiple of 32

[@abhishekkumarsrivastava7677]

After training using darknet on traffic sign dataset, it generates all weight files of approx. size 244MB. This size is not suitable for embedded applications. How can I reduce the weight size to less than 5 MB?

[@CodeWithAarohi]

convert your darknet model into tensorflow model and then to tflite. You will get a light weight model

[@abhishekkumarsrivastava7677]

@@CodeWithAarohi Will 244MB model be reduced to less than the 5MB model after conversion or do I need to modify my input image size, dataset size, etc? I am asking it because, in your second video, ur weight file size was 22MB but in my case weight file is 224MB so I am doubtful that tflite could do this much reduction in model size!

[@abhishekkumarsrivastava7677]

@@CodeWithAarohi how come the converted TensorFlow model is approx 2 MB and tflite model is 23 MB? It should have been the other way round, I guess! I wanted to use tflite on microcontrollers but this model size is pretty heavy. Please comment

[@CodeWithAarohi]

@@abhishekkumarsrivastava7677 tensorflow model should be in greater in size as compare to tflite model. 2nd you can reduce the size of tflite model by quantize.

[@abhishekkumarsrivastava7677]

@@CodeWithAarohi I have rechecked my process multiple times, where I am creating yolo v4 tiny model weights using Darknet and then converting to tflite model using instructions given in video. For some strange reason, my tflite model file is equivalent to the size of yolov4 tiny model size. Even after quantization for Float 16, tflite model size is greater than yolov4 tiny. Can you please direct me u where I might be going wrong or where should I look to resolve this issue?