A multi-panel scientific diagram of a computer vision pipeline, showing input images, DINO features, sparse point cloud, 3D Gaussians, rendered RGB, rendered features, and an optimization loop.

A clean academic computer vision pipeline figure in CVPR style, horizontal multi-panel layout on white background, with labeled panels (a) to (g). (a) Input: multiple RGB images of a scene from different viewpoints, arranged in a row, with small camera frustums. (b) DINO features: same images transformed into colorful semantic feature maps (smooth heatmaps, patch-based colors). (c) Structure-from-motion: sparse 3D point cloud with camera pyramids in 3D space. (d) 3D Gaussian splatting representation: many semi-transparent ellipsoids (gaussian blobs) in 3D, colored, representing both RGB and feature embeddings. (e) Rendered RGB: photorealistic image rendered from the gaussian scene. (f) Rendered feature map: smooth colorful embedding visualization corresponding to DINO features. (g) Optimization: diagram showing loss computation, arrows comparing rendered RGB to ground truth RGB and rendered features to feature maps, with a loop arrow going back to the 3D gaussians. Style: clean vector graphics, minimal text, thin arrows, consistent color coding (blue for RGB, green for features), soft gradients, professional scientific figure, high clarity, no clutter. Mehr sehen