Adapting Wavelet Compression to Human Motion Capture Clips Philippe Beaudoin 1 Pierre Poulin 1 Michiel van de Panne 2 1 Université de Montréal, LIGUM 2 University of British Columbia, Imager

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface A need for compression? Motion capture is very popular Motion capture rapidly produces huge collections of data Escalating cost of the memory hierarchy (ie. Martin Walker talk)  Lossy compression

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface What is a good compression? Depends on the application We aim for: –Small cache footprint –Access to subset of joint data –Accurate foot placement –Independent motion clips Best ratio may not be the target

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Which kind of compression? Joint correlation –2:1 up to 4:1 (PCA) Joint + temporal coherence –Cannot access individual signals –30:1 up to 35:1 [Arikan 06] Temporal coherence alone –35:1 (this work) –Access to subset of joint data –Low computational requirements

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Preliminary details… A pose is… –Root position (3 signals) –Euler angles of joints (59 signals) Motion is sampled at 120 hz No preprocessing or format conversion before compression

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Standard wavelet compression Cubic interpolating bi-orthogonal wavelet basis [Sweldens 98] Not specially targeted to motion capture

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Standard wavelet compression Wavelet transform 62 signals Keep the largest coefficients from all the transformed signals Yield vector w i (1 ≤ i ≤ 62) counting how many coefficients are kept for each signal

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Vector w i

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Optimized coefficient selection w i minimizes RMS error in the DOF Quality depends much more on positional distortion Optimally redistribute coefficients? –Too costly!

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Optimized coefficient selection Motion capture data is hierarchical Build vector m i that favors some signals more than others Fixed choice for m i ? Bad! –Depends on complexity of signals –Depends on the poses

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Start with m i = w i

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Randomly select i reduce m i

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Find optimal j increase m j Repeat…

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Inverse kinematics correction Problem: Noticeable sliding feet Change distortion metric? –Assumption breaks down, difficult to find a good m i Instead, add positional channels for the feet, use IK

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Inverse kinematics correction Signals encode difference between compressed position and true feet position Wavelet compress these signals independently

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Final details… Quantize to 16 bits Run-length encode 0s Optionally use lempel-ziv independently on each clip

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Results Tested on 1 sec. to 45 sec. clips Compression ≈ 300 ms/frame Decompression ≈ 30 μs/frame(no IK) ≈ 300 μs/frame(with IK)

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Video

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Conclusion Tractable coefficient search space adapted to motion capture data Fast decompression Access to subset of joints Independent clips 35:1 compression ratio

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Future Work Metric for perceived quality in a motion capture animation Explore large-scale redundancies (see our technical report) Level-of-detail streaming

Adapting Wavelet Compression to Human Motion Capture Clips Beaudoin, Poulin, van de Panne – Graphics Interface Questions?