Super pixel Guided Near to Far Field Learning

1
Super pixel Guided Near to Far Field Learning

• Soumya Ghosh

2
The Problem

• Autonomous Robot Navigation in unstructured
  environment.
• Distinguish between traversable ground and
  obstacles in the given image.
• Semantic Image analysis.

3
(No Transcript)
4
Solution

• Typically we rely on stereo data.
• Stereo labels are usually present for both
  classes in the near field.
• If abundant, they can be used for planning
  directly.

5
Near to Far Learning

• Classify the image far field into ground plane
  and obstacles.
• Near Field 10m
• Mid Field 10 -25m
• Far Field 25m and beyond.
• Near Field 60 of the image
• Far Field 20 of the image

6
Need for Near Far learning
7
Issues

• Extremely sparse stereo labels in far field.
• Objects appear smooth in the far field. Texture
  mapping from near to far is a bad idea.
• Scale issues.

8
Motivation behind the proposed solution

• We address the problem of sparse stereo labels.
• Poorly sampled sparse training data does not
  capture intra class variance.
• Make assumptions.

9
Proposed Solution

• Assume the first law of geography Everything
  is related to everything else but near things are
  more related than distant things to hold.
• We can then assume that the training points
  capture the intra-class variance in some
  neighborhood N defined around them.

10
Proposed Solution

• We restrict the classification process to only
  this neighborhood and deem everything else as
  unknown.

11
Algorithm

• Segment the far field into a collection of
  segments S.
• Find segments which neighbor the stereo ground
  plane labels. Sgnd.
• Find segments neighboring the stereo obstacle
  labels. Sobs.

12
Algorithm

• Compute color histogram over the regions stereo
  labels as ground plane (hgnd) and also the ones
  stereo labels as obstacles (hobs).
• Compute color histograms for all segments in Sgnd
  and Sobs.
• If the segments in Sgnd are less than some
  threshold dgnd away from hgnd label them as
  ground plane. Do the same for obstacles.

13
Algorithm

• If a segment belongs to both Sgnd and Sobs and
  lies within the dgnd and dobs range, then the
  segment is ambiguous and classified as unknown.

14
Illustration
15
Determining dgnd and dobs

• Over segment the mid-field region of the image
  frame.
• Compute the histogram distances between the mid
  field stereo ground plane and mid field segments
  labeled as ground by stereo. These are segments
  with a majority of one class of pixels. The
  smallest (largest) such distance is dgnd.
• dobs is computed similarly.

16
Experiments

• Compared with Stereo (baseline) and linear SVMs
  (trained only on mid-far field data).
• Segmentation algorithms tested
• Mean Shift segmentation
• Graph based Segmentation
• K- means clustering

17
Segmentation Algorithm Comparison

• Biggest challenge was to standardize the
  different segmentation parameters for a
  reasonable comparison.
• 3 tunable parameters for mean shift hs, hr and
  min size.
• 2 parameters for Graph based segmentation
  aggressive merging quotient and min size.
• 1 parameter for k-means. K which governs the
  size of clusters.

18
Efficient Graph Based Segmentation

• No (normalized) min cut computation and
  transitively no eigen values computation.
• Makes it significantly faster.
• Bottom up strategy.
• Components are merged together when the inter
  component distance is smaller than the intra
  component distance of at least one component.

19
Efficient Graph Based Segmentation contd.

• Internal Distance maximum weight in the MST of
  the component t.
• Inter Component Distance minimum weight between
  the two components.

20
Results

• Graph based segmentation and mean shift
  segmentation both capture similar structure in
  the images. Graph based segmentation is however
  faster.
• K- means leaves blocky artifacts, and produces
  noisy segmentation.

21
Comparison with other classification techniques.
22
Illustrative Performance
23
Conclusion

• Performs better than stereo and a SVM trained on
  similar sparse data.
• Robust performance.