cheind/kdmap

Automatically exported from code.google.com/p/kdmap

Hey!

I'm Chris. My research connects perception, robotics and machine learning.
To learn more, visit https://cheind.github.io/ and browse my repositories 👇

Latest Updates
_{2024/03/13      Just published a code demo on image stitching via analytic homographies https://github.com/cheind/image-stitch
2023/08/01      My findings on the 'Anoto micro-dot codec' got published alongside with code https://github.com/cheind/py-microdots
2022/12/01      Added pure PyTorch implementation of Instant NGP for accelerated learning of NeRFs https://github.com/cheind/pure-torch-ngp}

kdmap's Issues

Trivial splits might cause endless recursion

What steps will reproduce the problem?

1. Create a node with split bounds bigger than the contained elements (as it 
can happy when using midpoint splitting)
2. Add two or more points with equal coordinates
3. Try to split.

What is the expected output? What do you see instead?

The ISplitDimensionPolicy should throw a DegenerateDatasetException. Instead 
an endless recursion might occur.

Original issue reported on code.google.com by [email protected] on 15 Jun 2009 at 8:41

Implement dictionary based on kd-tree

The class on top of kd-tree driving the dictionary access has not yet been 
implemented.

Original issue reported on code.google.com by [email protected] on 27 Jun 2009 at 4:27

Major performance issue

    private void TrySplit(KdNode<T> target) {
      try {
        _subdiv_policy.Split(target);
        target.Vectors = null;
        TrySplit(target.Left);
        TrySplit(target.Right);
      } catch (Subdivision.SubdivisionException) {}
    }

This implementation is bad.

if (target.Vectors.Count <= this.MaximumBucketSize)
        throw new BucketSizeException();

Line above should not be exceptional state, exceptions are very expensive and 
are not created for program flow control in C#. This is decreasing debugging 
speed enormously, and run speed also, but not as much. So if node is not big 
enough for splitting this should be handled by function return value or 
something, in any case not exception. 

Original issue reported on code.google.com by [email protected] on 29 Nov 2011 at 11:42

Documentation of Algorithms and Datastructures

Reference publications of used algorithms where appropriate

Original issue reported on code.google.com by [email protected] on 27 Jun 2009 at 9:17

Provide approximate (epsilon-based) nearest neighbor queries.

Currently epsilon-based approximate nearest neighbor queries are not supported.

Original issue reported on code.google.com by [email protected] on 28 Jun 2009 at 7:47

Simplify KdTree<T>.FindInSortedOrder

looks too clouded

Original issue reported on code.google.com by [email protected] on 8 Jun 2009 at 7:16

Provide replacements for constructor overloading

Scripting languages like for example Ruby cannot handle constructor 
overloads. Instead, provide static methods that create such objects.

Original issue reported on code.google.com by [email protected] on 21 Jun 2009 at 1:04

NUnit 2.5 is required to run unit test

Assert.Throws is only available on latest NUnit 2.5 which in turn seems to
be unstable under Mono

Original issue reported on code.google.com by [email protected] on 21 Jun 2009 at 4:17

Calling KdTree<T>.Clear will render all

What steps will reproduce the problem?
1. Create a KdTree
2. Create a ClosestLeafSearch on the KdTree
3. Clear the KdTree
4. Search

Original issue reported on code.google.com by [email protected] on 28 Jun 2009 at 12:58

RenderTree misses OptProperties.cs

file seems to be missing in trunk

Original issue reported on code.google.com by [email protected] on 21 Jun 2009 at 5:26

Handle equality other than based on coordinates

Currently ExactSearch and others determine equality based on coordinates. 
Other class may want to specify equality explicitly as equality on 
coordinates and something else.

Original issue reported on code.google.com by [email protected] on 28 Jun 2009 at 12:04

KdMap.TryGetValue seems to be in-efficient

class Numbered uses a bunch of enumerator to reach its goals. Consider an
implementation of TryGet directly at ExactSearch

Original issue reported on code.google.com by [email protected] on 28 Jun 2009 at 2:28

Rendering.RenderTreeCairo.Render always takes KdNode<IVector>

How can one render a tree of KdNode<Custom>?

p = Accelerators::Subdivision::SubdivisionPolicyConnector.new(1)
# Create a new kd-tree in two-dimensional space
t = Accelerators::KdTree[Accelerators::Vector].new(2, p)

rendering t will currently not work because the type argument is Vector.

Original issue reported on code.google.com by [email protected] on 21 Jun 2009 at 7:43

Consider correct? implementation of ICollection.IsReadOnly property

by now the kd-tree, kd-map just return false.

Original issue reported on code.google.com by [email protected] on 28 Jun 2009 at 2:08

Unit tests for ICollection interface of KdTree is missing

add tests.

Original issue reported on code.google.com by [email protected] on 27 Jun 2009 at 4:28

Make subdivision more flexible

Instead of repeating code in subdivisions (choosing axis, choosing
splitvalue), refactor to common interfaces. Use a connector to bring the
interfaces together and thus provide an implementation for ISubdivisionPolicy.

Original issue reported on code.google.com by [email protected] on 9 Jun 2009 at 4:18

_trivial_resolver null

 public SubdivisionPolicyConnector() {
      _max_bucket_size = 25;
      _dim_selector = new Accelerators.Subdivision.PeriodicAxisSelector();
      _loc_selector = new Accelerators.Subdivision.MidpointSelector();
    }

 _trivial_resolver = new NoOperationResolver() is not set in default ctor so you get null exception later.

Original issue reported on code.google.com by [email protected] on 29 Nov 2011 at 11:26

Provide support for specifying the dimension of vector, kd-tree, aabb etc. after instantiation

Currently all constructors require at least a dimension parameter. 
Serialization might not be happy about this.

Original issue reported on code.google.com by [email protected] on 27 Jun 2009 at 4:32

Stress-tests and benchmarks missing

Need some timing/memory information

Original issue reported on code.google.com by [email protected] on 22 Jun 2009 at 5:35

Test Accelerators.Searches.ExactSearch<T>

Has been added without a test

Original issue reported on code.google.com by [email protected] on 28 Jun 2009 at 2:00

VectorOperations should use 'ref'

As a struct might implement the IVector interface it is necessary to flag all 
by reference parameters with the ref keyword. See 
Accelerators.VectorOperations for example.

Original issue reported on code.google.com by [email protected] on 21 Jun 2009 at 6:04

Implement serialization suppert

Serialization support for kd-tree and driver classes is required.

Original issue reported on code.google.com by [email protected] on 27 Jun 2009 at 4:29