Ahsen Khaliq’s Post

I have color vision tested twice a year and still have 100% accuracy. I do find how others see color to be very interesting. Including how it may deviate from the rating methods. https://lnkd.in/gxNGJ864

Proud to share that two of our research papers have been accepted at ECCV 2024! 🎉 📄 Paper 1: "HandDAGT: Denoising Adaptive Graph Transformer for 3D Hand Pose Estimation" 👋 This paper addresses the challenges in 3D hand pose estimation, particularly self-occlusion and intra-occlusion. HandDAGT employs a novel transformer structure and attention mechanism to enhance robustness and accuracy, showing significant improvements on four challenging datasets. This work builds on our previous successes on 3D vision tasks, including ECCV22 oral and CVPR24 highlight papers. Explore our work, and access our codes & pre-trained models here: https://lnkd.in/gYQpDdU8. 📄 Paper 2: "CRAD: Continuous Representation for Anomaly Detection" 🔍 We introduce CRAD, a method that uses continuous memory representation to improve unsupervised anomaly detection, especially for multiple object classes. Special thanks to my collaborators Eunbyung Park and Simon S. Woo. Notably, our first author, Joo Chan Lee, adapted the grid concept from his work on 3D modeling, which was recognized with ICLR 24 spotlight and CVPR 24 highlight papers, to the anomaly detection domain in CRAD. Check out our project page for more details: https://lnkd.in/g_xFskTP.

Exciting new research alert! Check out this insightful blog post on curve-based neural style transfer. The study introduces a novel parametric framework tailored for curve-based design sketches, addressing the challenges commonly encountered by traditional neural style transfer methods. By combining shape-editing rules and efficient curve-to-pixel conversion techniques, the research enhances the role of VGG19 on ImageNet-Sketch, paving the way for precise style extraction and potentially revolutionizing design articulation in product design. Read the full post here: https://bit.ly/42anNC3.

Depth Anything V2 🔔🔔🔔 Depth Anything V2 introduces significant improvements in monocular depth estimation by using synthetic images, scaling up the teacher model’s capacity, and employing large-scale pseudo-labeled real images. Compared to its predecessor and the latest models, it is more efficient and accurate, offering models ranging from 25M to 1.3B parameters for various scenarios. Additionally, it includes a versatile evaluation benchmark with precise annotations and diverse scenes to enhance future research. project: https://lnkd.in/dAq-Eck5 code: https://lnkd.in/dWTteRHM demo: https://lnkd.in/dHdsnt-j paper: https://lnkd.in/dkY2tfXd #DepthAnythingV2 #MonocularDepthEstimation #SyntheticImages #AIResearch #MachineLearning #EfficientModels #DepthPrediction #ComputerVision #Benchmarking #ModelTraining

While exploring the working of different face recognition models, I realized that classic SoftMax loss function has some major drawbacks. While Softmax is a novel approach for closed set classification problems, but it is not suitable for open set face recognition problem. Reasons: 1. Inter class discrimination: SoftMax focuses on finding a hyperplane that separates classes but does not constrain how close the datapoints of classes are around their mean in the feature space, due to which even if the faces are all correctly classified, since there may be pose and expression variations, light conditions or some other variations, the model may still place the same individual’s images at various distance from each other in the feature space.  2. Intra class discrimination: while SoftMax loss focuses on classifying images correctly but it does not focus on maximizing the distance between classes. This means that while two people may be classified correctly, their representation or position in the feature space may be closer than expected, this induces a high chance of misclassification in borderline cases. There are many modifications proposed to address these problems, the most novel and practical method is Arcface (Additive Angular Margin Loss). As per my understanding of the 1801.07698.pdf (arxiv.org), the Arcface modifies the existing SoftMax function to induce angular margin. Angular margin: Arcface introduces an angular margin m into the SoftMax loss function, which effectively increases the angular distance between the different classes in the feature space. When this margin is added to the angle between θyi between the feature vector and the correct class weight vector, it forces the decision boundaries between classes to be more precise. Normalization: It normalizes both the input face features and the weight vectors representing each class. This forces predictions to rely only on the angle between them. Increasing Intra-Class Compactness: By normalizing features and weights, ArcFace constrains representations of the same class to lie closer together in a hyper spherical space. This induces a tighter clustering, enhancing intra-class discrimination and allowing the model to better accommodate natural variations. #machinelearning #facerecognition #neuralnetworks

More website (emergentmath.com) redesign & resource-share. Here's one that I really like but haven't found the right "home" for other than on my blog. I'd like to publish it in a physical journal/magazine/book so but I'm waiting for the right place. A CCSS Standards of Mathematical Practice aligned rubric. Use this rubric to assess students and give meaningful feedback on the 8 standards of mathematical practice. #mathematics #matheducation #commoncore #mathematicalpractices #assessment https://lnkd.in/dq7MCa7

#gameengine #machinelearning #probability Let X1 , X2, X3, ..., Xn be a random sample from a normal distribution N(θ,1) prove X̄∼N(θ,1/n), in the book, author just concludes this, no proof. I went an extra mile, proved it! (proof is based on another proof: MGF of Y=aX+b is given by MY(s)=e^sb MX(as)) the equation below(center limit theorem) is the most useful formula so far, which can be proved the same as the proof above!!! Q = (X̄-θ)/ (σ/√n) ~ N(0,1) first time derived/proved distribution of a random variable using moment generating function.  this MGF is something really interesting here, it's the same as Laplace transform(with a flip of sign) and, it connects to power/Taylor series, and it tells moments, which is a concept in physics. learning is fun, right? I would not dream I can come to this point, and yes, I am here, close to the ending of the course :) P.S. if anyone needs intuition about convolution: PDF of sum of two random variables is the #convolution of the two individual PDFs.

I am excited to share my article on calculating gain and phase shift using differential equations! As part of a recent assignment, I delved into the intricate world of differential equations, and the calculation of gain and phase shift. This exploration wasn't just about solving equations, it was a dive into understanding how these concepts apply to real-world scenarios like audio and control systems. Using a workbook and other resources I was able to: - Understand and successfully calculate the steady-state solution. - Calculate the gain and phase shift, bridging the gap between theoretical math and practical application. - Use Sage math to create and graphically analyze the input system compared to the output system. Helping to visualize the transient and the steady-state solutions as well as the gain and phase shift. I spent a lot of time trying to understand and grasp the process of finding the gain and phase shift within the steady-state solution. The toughest challenge being the transition between real and imaginary numbers in Cartesian and Polar systems. This paper provides a deeper understanding of differential equations, highlighting their critical role in audio and control systems, although the focus remains on the mathematical foundation rather than specific applications. Saint Michael's College #Mathematics #DifferentialEquations #GainAndPhaseShift #SteadyStateSolution #MathInRealLife #ControlSystems #AudioEngineering

Thrilled to share that our paper "MMM: Generative Masked Motion Model" has been accepted at CVPR 2024 as 🎉a highlight poster🎉. In this work, we introduce the Generative Masked Motion Model for the Text-to-Motion task, which simultaneously achieves high-fidelity, high-speed motion generation, and editability. MMM outperforms diffusion models and autoregressive approaches, with superior FID scores of 0.08 and 0.429 for the HumanML3D and KIT-ML datasets, while offering advanced editing features such as body-part modification, motion in-betweening, and the synthesis of long motion sequences. In addition, MMM is two orders of magnitude faster on a single mid-range GPU than editable motion diffusion models. 📑 Paper: https://lnkd.in/eBi6r8h4 🕸 Project Page: https://lnkd.in/eziKRjR3 👨💻 Code: https://lnkd.in/eV24MsfE This work wouldn't have been possible without help and advice from Pu Wang Chen Chen Minwoo Jake Lee

NCERT 10th Mathematics Exercise 7.4 Topic: 7 Coordinate geometry To learn all the problems of the above exercise systematically in an effortless way, Click on the following link: https://lnkd.in/d4hU-U5P Need help with math? We're here to help! Our resources include NCERT textbooks, lessons on Coordinate geometry, and more. Join our community of students and teachers today! #mathhelp #NCERT #studentsuccess #Coordinategeometry #education #learning #students #teachers #math #educationpolicy