Stanford OnlineStanford AA274A Principles of Robotic Autonomy | Autumn 2019 | Modern Computer Vision
Stanford Online
Course feedback 0:05
The instructor responds to questions about sections, homework, and grading. Sections are meant to complement lectures with hands-on robotic software engineering that is hard to teach from slides. He acknowledges connectivity and installation problems, says they are being fixed, and notes that homework feedback before the midterm was delayed because auto-grading is new and slow to set up. Midterm grades should be released by the end of the day or by Sunday.
Vision before 2012 8:02
The lecture then turns to modern robotic perception. Boris frames it as a look at computer vision before and after 2012. Earlier work relied on engineered features such as edge, corner, and blob detection, plus descriptors like SIFT, SURF, and ORB. The aim was to squeeze useful information from raw pixels by hand, using methods built around structure and strong assumptions.
Semantic segmentation 16:31
Semantic segmentation means classifying every observed point in a scene. A simple way to do it is K-means clustering. Each pixel is mapped into a five-dimensional feature space made from X, Y, and RGB, so nearby pixels with similar color can be grouped together. With a small K, you get coarse classes like island, shallow water, and deep water. With a larger K, the picture becomes finer, but too large a K is expensive and loses meaning.
Graph methods and SLAM 25:00
Graph-based methods use the image grid more directly. They treat pixels as random variables in a probabilistic graphical model and connect each pixel to its neighbors, or sometimes to a richer set of related pixels. A first classifier gives a noisy guess, especially near edges, and the graph model sharpens it through global optimization. The same geometry-first approach also matters in SLAM, where you must locate yourself while mapping the scene.
ORB SLAM and keypoints 32:00
ORB-SLAM detects stable key points first, then estimates rotation and translation between frames. The green dots mark points chosen because they stay easy to find under light change, shadow, and rotation. That makes localization strong. It does not make a dense map, because only a small share of pixels are used, so the result is sparse rather than full.
Dense mapping and loop closure 36:33
DSO, or direct sparse odometry, uses denser image features like edges and keeps a more detailed map. It still has drift, especially when you return to a place you already saw. Loop closure fixes that by noticing the repeat visit and correcting past trajectories and the map so they match again.
Convolutional networks 50:30
The lecture turns to convolutional neural networks, or CNNs, as the method that cut ImageNet error from 26% to 16%. A CNN is a predictor built from stacked, trainable filters. Early layers learn simple edge detectors, including color edges. Middle layers combine edges into corners, circles, grids, and slopes.
From filters to prediction 56:00
A demo shows a seven passing through the layers of a CNN. The first filters respond to horizontal and diagonal edges, while higher layers become harder to interpret, which is why these models are called black boxes. The network ends with a fully connected classifier, can read other digits, and later reached superhuman ImageNet accuracy, though it is slow.
CNNs reshape vision 1:05:02
Post-2012 vision methods increasingly use one convolutional neural network to do several jobs at once. The network can find regions of interest, classify them, and learn bounding-box regression for location. The same approach now handles object detection, tracking, motion prediction, and even image generation, with the loss tied to pixel-level errors. Older graph methods and clustering still appear, but mostly as support tools. In robotics, CNNs now help with SLAM and depth prediction, while sensor inputs from cameras, lidar, GPS, IMU, and wheel encoders are combined as extra channels for better results.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
