대충공부한거적어두는블로그

https://arxiv.org/abs/2410.15778https://github.com/shengliu66/VTI GitHub - shengliu66/VTI: Code for Reducing Hallucinations in Vision-Language Models via Latent Space SteeringCode for Reducing Hallucinations in Vision-Language Models via Latent Space Steering - shengliu66/VTIgithub.com깃헙 레포는 글 작성일인 오늘 업데이트된 따끈따끈한 논문..인데 어쩌다보니 읽어봄 Abstraction + IntroductionLVLM의 환각은 시각적 입력과 텍스트 출력 간의 불일치로 인해 발생하는..

LLaVA OneVision을 fine-tune 해보고 있는데, LVLM의 mm_vision_tower (vision encoder)로 siglip을 사용중이어서 이참에 정리 겸 공부해볼까 싶어 남긴다 이미지에서 시각 정보를 vision feature vector로 추출하고 그것을 text input과 동일한 embedding 공간에 투영하는 과정을 거쳐 text decoder (여기선 LLM이겠죠)의 인풋으로 들어간다.. encoder 모델CLIP (Contrastive Language-Image Pre-training)* CLIP은 이미지와 텍스트를 같이 학습하여 두 가지의 상호 연관성을 강화하는 contrastive learning 기법을 사용* 이미지 인코더(ViT)와 텍스트 인코더(BERT)를 ..

2D image points are OK which we can easily find from the image. (These image points are locations where two black squares touch each other in chess boards) we need to know 3D points(X,Y,Z) values. But for simplicity, we can say chess board was kept stationary at XY plane, (so Z=0 always) and camera was moved accordingly. ⇒ Now for X,Y values, we can simply pass points as (0,0), (1,0), (2,0) … (s..

1. 이미지 열기cv.imread : 첫 인자는 파일경로IMREAD_COLOR loads the image in the BGR 8-bit format. This is the default that is used here.IMREAD_UNCHANGED loads the image as is (including the alpha channel if present)IMREAD_GRAYSCALE loads the image as an intensity onecv.imshow : 화면에 뵈기cv.waitKey(0) : wait for user input in ms (0 ⇒ forever)cv.imwrite : image is written to a file path 2. 비디오 캡쳐해서 보여주기 및 저장하기 (얘..

https://github.com/opencv/opencv GitHub - opencv/opencv: Open Source Computer Vision LibraryOpen Source Computer Vision Library. Contribute to opencv/opencv development by creating an account on GitHub.github.comhttps://docs.opencv.org/4.x/d6/d00/tutorial_py_root.html OpenCV: OpenCV-Python TutorialsCore Operations In this section you will learn basic operations on image like pixel editing, geome..