Comparing NARF and SIFT Key Point Extraction Algorithms Chris Kaffine Second Annual MIT PRIMES Conference, May 20th, 2012

Range Sensors Purpose: collect distance information Advantage over cameras: 3D Methods: Stereo Imagery LiDAR Structured Light

Representing Range Data Point Clouds: 3D-coordinates Geometrically understandable Range Images: 2D-image with pixel values representing depth Similar to sensor functioning Allows border extraction

Correspondences Goal: Find points in two images which are equivalent With matched points, differences between images can be calculated

Key Points and Descriptors Find correspondences in two steps: find key points, calculate descriptors Key Points- Distinguishable, stable locations in a scene Descriptors- Numerical description of a point and its underlying surface Points with similar descriptors are correspondences

NARF Normally Aligned Radial Features Uses range images Uses borders and change in distance (pixel) values to identify key points Key points are invariant to scale, susceptible to camera orientation Support Size: indicates how detailed the search should be

SIFT Scale Invariant Feature Transform Uses point clouds Finds key points that are invariant to scale Utilizes full, 3D geometry Scale Size: indicates how close to “zoom in”

Evaluating the Algorithms Use data with known sensor location Within chronologically adjacent frames, search for nearby key points Points within a certain distance are considered true matches Count number of frames each point lasts for Repeat, using different algorithms with different parameter values and different distance thresholds

Evaluating the Algorithms Metrics for evaluation: Number of key points identified Persistence/Stability of key points Density of key points, with relation to distance threshold Due to limitations in persistence algorithm, two persistence metrics were used: Measure 1:Average persistence of all key points Measure 2: Number of key points with persistence greater than 1

Results- Measures of Success Measure 1: Smoother, NARF exceeds SIFT in parts Overall, similar trends, though distinct metrics

Results- Measures of Success Measure 1: Smoother, NARF exceeds SIFT in parts Overall, similar trends so overestimation most likely did not have a strong effect

Results- Measures of Success At low parameter values, SIFT key point numbers and density rise dramatically, NARF values rise steadily Indicates that as parameter values decrease, superfluous key points are detected

Results Best parameter values for each algorithm displayed Metric used: #key points * persistence / density SIFT almost always superior Scale size .07 better in general, 0.1 possibly better in some cases

Acknowledgements MIT PRIMES Professor Seth Teller Jon Brookshire – Mentor