What are the most important computer vision projects to include in your portfolio?

Deep learning methods, especially convolutional neural networks (CNNs), have achieved remarkable results in image classification. However, deep learning methods also have some drawbacks. First, deep learning methods require a large amount of labeled data to train the models, which can be costly and time-consuming to obtain. Second, deep learning methods are often sensitive to variations in the input data, such as noise, occlusion, rotation, scale, etc. Third, deep learning methods are often opaque and hard to interpret, meaning that they do not provide explanations for their decisions or errors. However, this drawback might be overcome using explainable AI (XAI) methods.

I think computer vision projects is the most impressive so, I'm using image classification and computer vision projects in my portfolio.

One that might be a useful way to showcase your skills is facial sentiment analysis at various sporting events that you can validate with actual events in the game. Many games are recorded, and you can mark the changing faces in the crowd informing your metrics against particular plays in the game. Or, for something more political, perhaps do a facial sentiment analysis of political media commentators towards a particular politician facing an election, and see if a particular sentiment is associated with the election result. Or, you can study the faces in the crowds of political election rallies. They might be good predictors of results and a computer can be used to build a model for the future.

Computer Vision AI is the widely used in Vision-based attentional control and one of the most critical technology to make AI accessible. Dr Peter Scott-Morgan - the British scientist who became ‘world’s first cyborg’ used computer vision extensively in his eye-tracking technology that enabled him to control multiple computers for his body movements. Amazon Palm One is revolutionising the payment and identity landscape using computing vison and images from our palm data to generate every individual's unique palm signature and building our identity profile which is more accurate than our Iris based biometrics.

Image generation is something to be very careful with from the perspective of copyright, privacy, and ethics. And, just because you generate an image this year that is in a bit of a legal grey area, say, a celebrity lookalike, and no one says anything, you might still be sued in the future when you are least expecting it.

DONT WASTE TIME ON GAN or VAE, START WITH DIFFUSION! When we have more powerful models like Diffusion, both Latent and Pixel based, it doesn't make sense for a practitioner to spend time on GANs, and VAEs anymore. Instead working on enhancing the output with the help of ControlNet and LoRA should be the focus for anyone working on applied Image Generation.

These can be useful but do not allow humans to be completely removed from oversight, and, matters that still depend on perfect image recognition, like law enforcement, require flawless and continuous model training. On the other hand, if you are making a corporate product, I highly recommend involving your users and testers in the perfection of the model, versus avoiding the assumption that your model is always in beta mode and irregularly, opaquely, and close-mindedly building it better.

we can also use Face Recognition - Face detection using HOG/SVM and recognition with FaceNet embeddings. Valuable real-world application. Style Transfer - Implement neural style transfer to recreate artistic styles. Shows ability to manipulate representations. Visual Question Answering - Answer text questions about images; involves image understanding and NLP. 3D Computer Vision - Projects using 3D data like point clouds or meshes for perception and reconstruction.