Common Pitfalls to Avoid in Object Detection Datasets - Object Detection Challenges & Solutions
HTML-код
- Опубликовано: 18 июн 2024
- Learn about the best practices in creating high-quality datasets for Object Detection. “Data is the new Oil” - Unrefined and unpolished data will only result in a “GIGO” (Garbage In, Garbage Out) system!
Many Deep Learning practitioners ignore the importance of data quality while building the model and keep iterating over model building instead of improving their data. Here we discuss ideas on how to analyze your dataset and common pitfalls while creating the dataset. We also talk about how checking your data gives you insights into the quality of your dataset as well as tips on how to improve the data and, eventually, the model performance.
We take an example of a freely available public dataset to discuss the various issues that you may encounter while solving an Object Detection problem.
⭐️ Time Stamps ⭐️
0:00-00:22: Motivation
00:22-1:15: The Dataset
1:15-3:03: Analyzing the Dataset
3:03-4:29: Tip: Visualize the Dataset
4:29-6:14: Understanding the classes
6:14-7:54: Pitfall: Oversampling frames from a video
7:54-11:36: Data Variance vs Data Size
11:36-11:57: Tip: Compare Training and Validation Set
11:57-14:35: Training Validation Overlap
14:35-16:01: Tip: Check Data Statistics
16:01-17:01: Pitfall: Class Imbalance
17:01-20:33: Visualize Data Annotations
20:33-21:34: Pitfall: Miscalssified or Incorrect Labels
21:34-27:03: Pitfall: Missing / Wrong Labels
27:03-29:22 : Pitfall: inconsistent labels
29:22-31:11 : Summary
🖥️ On our blog - learnopencv.com we also share tutorials and code on topics like Image Processing, Image Classification, Object Detection, Face Detection, Face Recognition, YOLO, Segmentation, Pose Estimation, and many more using OpenCV(Python/C++), PyTorch, and TensorFlow.
🤖 Learn from the experts on AI: Computer Vision and AI Courses
YOU have an opportunity to join the over 5300+ (and counting) researchers, engineers, and students that have benefited from these courses and take your knowledge of computer vision, AI, and deep learning to the next level.🤖
opencv.org/courses
#️⃣ Social Media #️⃣
📝 Linkedin: / satyamall. .
📱 Twitter: / learnopencv
🔊 Facebook: profile.php?...
📸 Instagram: / learnopencv
🔗 Reddit: / spmallick
🔖Hashtags🔖
#AI #machinelearning #objectdetection #deeplearning #computervision #datasets #pitfalls #objecttracking #dataset #bestpractice
Get expert guidance, insider tips n tricks and Create stunning images, learn to fine tune diffusion models, advanced Image Editing techniques like In-Painting, Instruct Pix2Pix and many more. Gain the first-mover advantage with OpenCV Master AI Art Generation Course. Join our Kickstarter campaign now! bit.ly/3JYh7A6
Excellent video. I realized I made multiple mistakes during the first iteration of my training. I currently focusing on creating a better dataset which is more representative.
Good for you!
Time Stamps:
0:00-00:22: Motivation
00:22-1:15: The Dataset
1:15-3:03: Analyzing the Dataset
3:03-4:29: Tip: Visualize the Dataset
4:29-6:14: Understanding the classes
6:14-7:54: Pitfall: Oversampling frames from a video
7:54-11:36: Data Variance vs Data Size
11:36-11:57: Tip: Compare Training and Validation Set
11:57-14:35: Training Validation Overlap
14:35-16:01: Tip: Check Data Statistics
16:01-17:01: Pitfall: Class Imbalance
17:01-20:33: Visualize Data Annotations
20:33-21:34: Pitfall: Miscalssified or Incorrect Labels
21:34-27:03: Pitfall: Missing / Wrong Labels
27:03-29:22 : Pitfall: inconsistent labels
29:22:31:11 : Summary
Great video! While the background music isn't loud, to my ears, it is a little intrusive and not needed.
Noted!
Thanks Sir for this informative video. The content of this Video is pure gold.
I have been doing the Exploratory Data Analysis and Overlays for a while and many times people think it is a waste of time to go at such granular level to visually examine the data.
Now, I have your this video to prove my point.😊
Thanks Sir. 🙏
Thank u for the video! How should be prepared dataset for long or short objects passing on conveyor belt?
It was greatly helpful. Glad that you uploaded it!
Glad it was helpful!
hi ! thank you for this video, it's great. What software do you use to label ? Thanks
Hi Opencv, can you share the data stats code which is used in this example?
I like this video! It answered a lot of questions I had as a beginner. Thank you so much!
One question. This video is mainly about bounding box annotation. What about with key-point annotation? I am going to annotate mice in a cage, which means the objects are highly occluded. But I would like to use key-point annotation to detect their behaviour. What would be the best way to annotate to be consistent do you think?
We can use annotation tools such as imagelab, roboflow, etc for annotating keypoints
Thanks - great advice
I have one very urgent question. Meanwhile i was successfull in running yolo on my local gpu and training on it. But all the tutorials just show how to create a custom dataset with 1 or 2 classes. How would i add my custom datatset to an already existing like the coco one? Can you help?
In medical images, I used augmentation. Do you think that augmentation pollutes sets? I used 3 augs per image + I had some frames from the same video so I am going to change that (plus those where augmented)
No, augmentation does not "pollute" the dataset. Instead, it increases the data distribution and variation. This allows the model to learn more varied and complex features, and further leads to better convergence.
But, keep in mind that augmentations are "problem-specific" and there is no "one technique fits all" approach.
Thanks for informations
thanku sir
Glad you found it helpful!
Very well explained
Do you know about how to resolve issues of occluded objects? Can you provide some material to read on occluded objects.
@@sparklesmastiyo1642 What is the"issue"? How occluded objects should be treated depends on you. Do you want your detector to detect them? If so, you should ensure they are labeled and labeled consistently. Funny, as I'm writing this, Sateya is talking about occlusion. I'm on mobile so I can see the time stamp. It is just before the summary.
Thank you for the reply @cyberhard!
@@LearnOpenCV you're welcome.
Very nice sir. Superbh
Thanks a lot!
😊😊
▶ LINK TO YOLO MASTERCLASS PLAYLIST: ruclips.net/p/PLfYPZalDvZDLALsG9o-cjwNelh-oW9Xc4
Discover the magic of AI-powered art creation in our new Mastering AI Art Generation Course. Learn how to create stunning AI-generated images. Get expert guidance, insider tips & tricks for creating beautiful art using cutting-edge generative AI technology.
Join our Kickstarter campaign now! bit.ly/3JYh7A6
"GIGO" 😄
COme on man!! we know this data can only make someone "RUclips DataScientist", you need to have minimum 20000-40000 Images per label to build the model with 70+ accuracy THAT YOU CAN SELL!!! This data is only to impress your gf :)
noice