Using the shape context in cluttered scenes is unreliable.

The shape context descriptor alone is not powerful enough to yield reliable point correspondences in cluttered scenes.

- It is difficult to recover the scale parameter, since normalizing the radial distances by the man/median point distances will no longer work.
- Object and non-object points close to the object are hard to distinguish on the basis of their shape context alone.
- Points which are close to each other on the model shape are often matched to points which are far away from each other in the image.
- Shape contexts are not invariant under arbitrary affine transforms, but the log-polar binning ensures that, for small locally affine distortion due to pose change, intra-category variation etc., the change in the shape context is correspondingly small.
- The richness of the shape context makes it robust to noise and occlusion, but it is too complex and time-consuming.

**solutions:**

**RANSAC**: if some of the correspondences are correct, to identify outliers in the alignment phase of the iteration process using a robust estimation scheme. Outliers can then be excluded from the next shape context computation.
**Shapemes**: **vector quantization** on the shape context.
- Vector quantization involves clustering the vectors and then representing each vector by the index of the cluster that it belongs to.
- k-means clustering to obtain k shapemes: all of shape context from known set are considered as points in a d-dimensional space, so each know view is a collection of shapemes, each d bin shape context is quantized to its nearest shapeme and replaced by the shapeme label{1,…,k}.
- A known view is simplified into a histogram of shapeme frequencies. From s d-bin histogram to a single
** k-bin** histogram

Using Shapemes for Mid-level Vision

**How do we model context? **

- One way is to find a computational/operational definition for each
**mid-level visual cue**, such as continuity, convexity, or parallelism.
- An alternative is to use a generic “context” descriptor, such as shape context or geometric blur.

**Using edge orientation:**
- Only matching points with similar gradient orientation.

**Shape Context with Figural Continuity**
- The neighboring points on the model shape U should map to points on the target shape V which are also close to each other.
- The correspondences are the sum of shape context costs and the continuity cost.

### Like this:

Like Loading...

*Related*