홍석쓰 블로그

•What is the appropriate and differentiable norm when given data points with outliers? (median is not differentiable) •How do you define outliers? What is the probabilistic meaning of outliers? 첫 번째 질문의 답이 p-norm with high p였던 거 같아서 p-norm, 2-norm, frobenius norm을 찾아보다가 svd, eigenvalue decomposition을 찾아보다 왜 p-norm with high p인 지 아직도 이해를 못함. 일단 induced norm 인가 schatten norm인 지 모르겠음. https://chris..

09.18.2024Gradient Penalty라고도 불리나 봄. Applying a gradient penalty in GAN training (especially in WGAN-GP) helps stabilize the training process by enforcing the Lipschitz continuity condition. This regularizes the discriminator’s gradients, ensuring that they remain well-behaved, which in turn helps the generator receive more meaningful updates. It addresses critical issues like mode collapse, van..

Signal Processing in Computer Vision 복습. 카메라 팀에서는 이게 기본이구나. Radiometry, Stereo Depth 물어볼 줄 알았었는데. 아니었음. 이 짱짱 교수님 강의 정리해봄: https://www.youtube.com/watch?v=m_11ntjkn4k&list=PL2zRqk16wsdorCSZ5GWZQr1EMWXs2TDeu&index=8 Fourier Transform. Do you know Fourier Transform? What are the properties of Fourier Transform? If you have a periodic function, it can be expressed as a sum of sinusoids without loss ..

Stereo Depth Estimation 복습하다 feature matching파트에서 pixel level matching이 힘든 이유로 non-lambertian 인 경우가 있어서 lambertian과 specular reflection을 복습해봤다. (foreshorteningd, noise도 있는데 foreshorening은 region matching이나 object matching에서 더 문제 아닌가 싶음) 그리고 NeRF가 왜 좋은지, 즉 어떻게 non-lambertian refelecting surface를 modeling할 수 있는 지 공부하겠다. Stereo Depth Estimation: 1. Calibrate cameras: get extrinsics and intrinsics -..

rasterization code 쓸 때랑, sparse conv api 이해/설치하려고 할 때, bundle adjustment를 cuda로 짤수 있지 않을까 생각해봤을 때 (결국 안함.) CUDA를 접한 적이 있고, 여전히 솔직히 잘 모르겠어서 처음부터 정리해보려고 함. (02/18/2024) 아니 직접적으로는 웨이모 인터뷰 질문을 글래스도어에서 찾아보다가 What does thread mean for GPU? 라는 걸 보고, 흠 뭐라고 대답해야하지 라고 생각하다 이 영상 보고 정리해보기로 함. Intro to CUDA (part 1) https://www.youtube.com/watch?v=4APkMJdiudU - GPU > CPU: tremendous computational throughput ..

virtual viewpoint를 활용한 ordinal loss https://arxiv.org/pdf/2304.11970.pdf equation (7) 수식이 뭔가 잘못되어서 찾아보니 그냥 틀린거임. https://arxiv.org/pdf/1805.04095.pdf 실제 구현은 이러함 https://github.com/zerchen/gSDF/blob/master/common/loss/ordinal_loss.py#L145 참고: https://www.vedantu.com/question-answer/does-a-dot-product-equal-to-one-mean-class-11-physics-cbse-610cb4854d5f1d00a97380e8 What does a dot product equal to..

Download large google drive files Gdown is too unreliable - sudden stop after an hour, "too frequent xxxx", ... This seems to work; (Update 02/10/2024) This also fails since the token expires in an hour :( https://www.quora.com/How-do-I-download-a-very-large-file-from-Google-Drive How do I download a very large file from Google Drive? Answer (1 of 14): If you're looking for a quick and easy way ..

What is field of view? https://www.digitalcameraworld.com/tutorials/photography-cheat-sheet-what-is-field-of-view-fov Photography cheat sheet: What is Field of View (FoV)? Understand Field of view in photography and how it relates to the focal length of the lens and the sensor size – the diagram below shows you more www.digitalcameraworld.com FOV, focal length, sensor size https://www.e-consyste..