cs223b assignment 1 recap. lots of solutions! 37 groups many different approaches let’s take a...

Post on 20-Dec-2015

213 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

CS223BAssignment 1 Recap

Lots of Solutions!

• 37 Groups

• Many different approaches

• Let’s take a peek at all 37 results on one image from the test set

Possible Things to Look For

• Symmetry

• Color

• Shadow

• Corners

• Edges

• Texture

• Taillights

Road Finding Tactic 1

• Run adaptive flood fill on the image starting at the pixels near the bottom

• Everything that gets filled in is considered road

• Holes inside the road are likely car candidates

Road Finding Tactic 2

• Run an edge detector (such as Prewitt or Canny) and a Hough transform, and then look for lines with the right approximate angle and position.

Road Finding Tactic 3

• Sample the color from some pixels directly in front of car (bottom of image)

• Generate a difference map using this color

• Run an edge detector on this difference map, followed by a Hough transform

Road Finding Tactic 4

• Assume the road is in pretty much the same place in every image and hard code it in

Edge Approach 1

• Detect edges with Canny detector

• Dilate image to join nearby edges

• Fill in closed regions

Edge Approach 2

• Custom edge mask that considers vertical and horizontal edges in likely configurations

Shadow Finding

• Use a mask that enhances horizontal edges

• Look for several adjacent rows of dark pixels in the road region

Corner Approach

• Compute strong corners

• Cluster them with kmeans

• Find the densest clusters and compute their convex hulls

Texture Approach

• Look for regions of high frequency– Find density of edges in rectangular regions– Mask or sum the derivative across rectangular

regions– Take FFT of image and set magnitudes to 1,

leaving only phase info; inverse FFT then enhances portions with many frequencies

Taillight Detection

• Transforming the color space can help spot the reddest pixels

• Groups of these pixels are good candidates for taillights

• If there’s a horizontal pair of taillights, even better

• Or, use a mask with the pair built in

Going from Pixels to Regions

• Count the number of “car pixels” in square regions; if above a threshold, consider it a car region

Statistical Classifier

More stages mean many fewer false positives (and only slightly fewer missed cars)

Multiple Classifiers

Filtering Classifier Output Tactics

• Check number of corners in the rectangle.If too low, reject.

• Check position and size of rectangle. If it doesn’t make sense given perspective, then reject.

• Run an edge detector on the rectangular region and then sum the pixels inside. If below threshold, reject.

• Look at the average derivative of image intensity. If below threshold, reject.

Consider Geometry of Scene

Combining Results from Different Detectors

• Many of the best solutions used some weighted combination of several heuristics to verify the strength of hypotheses

Building a Better Bounding Box

• Get the strong corners in the box, and then trace the convex hull of these corners to eliminate some of the road.

• Alternatively, make a 2D or 3D model of car shape based on image position and replace rectangle with appropriate polygon.

• Or, segment the rectangle by colors and keep only the inner segments.

Common Mistakes

• Strange ways of drawing car regions

• Don’t use anti-aliased lines or flood fill, just draw filled polygons with cvPolyLine

Common Mistakes

• Stray Pixels

• Could be fixed with image close operation

top related