HW3

Any Copying from the work of another person is a violation of Carnegie Mellon University Policy on Academic Integrity.

To use

• Navigate to data and unzip all the data files. (I used up my LFS bandwidth so ...)

  • carseq.npy and girlseq.npy are datasets with purely translational objects (hence using a pure translation warp function in my Lucas Kanade tracker)
  • antseq.npy and aerialseq.npy are datasets with aerial views of moving objects from a non-stationary camera (hence using an affine warp function in my Lucas Kanade tracker)

• Run testCarSequence.py or testGirlSequence.py to see how Lucas Kanade (vanilla forward additive approach) works on carseq.npy and girlseq.npy with naive template update

• Run testCarSequenceWithTemplateCorrection.py or testGirlSequenceWithTemplateCorrection.py to see how Lucas Kanade (vanilla forward additive approach) works on carseq.npy and girlseq.npy with template correction

  • Different template update approaches are discussed at section 2.1 in Iain Matthews et al.

• Run testAntSequence.py or testAerialSequence.py to see how Lucas Kanade tracker works on antseq.npy and aerialseq.npy

  • Modify SubtractDominantMotion.py to choose Lucas Kanade Forward Additive approach or Lucas Kanade Inverse Compositional approach (a faster variant of LK) as the tracker
  • Modify SubtractDominantMotion.py to choose different combinations of erosion and dilation to visualize better results

Results

• Lucas Kanade w/ Forward Additive approach on carseq.npy
  • Baseline with naive template update in blue. Template correction in red.

• Lucas Kanade w/ Forward Additive approach on girlseq.npy
  • Baseline with naive template update in blue. Template correction in red.

• Lucas Kanade w/ Forward Additive approach or Inverse Compositional approach on antseq.npy

• Lucas Kanade w/ Forward Additive approach or Inverse Compositional approach on aerialseq.npy