Coder Evaluation

The 3D spatio-temporal orientation trees coupled with powerful SPIHT sorting and refinement, [6], renders 3D-SPIHT video coder, [8], so efficient that it provides performance superior to that of MPEG-2 and comparable to that of H.263 (which are generally used nowadays) with minimal system complexity. That is, 3D-SPIHT is one wavelet-based coder that exhibits state-of-the-art compression performance.

The comparative performance of the 3D-SPIHT with motion compensated temporal filtering (MC 3D-SPIHT) and the proposed coder (REVIC) without motion filtering, is tested on the data set of sequences of moving targets. In our problem the transmission of parameters is critical at extremely low bit rates. If motion compensated temporal filtering would be applied we will need to code motion vectors. Thus a key feature of the proposed video coder was to avoid the use of motion compensated temporal filtering, and consequently, motion vector components do not need to be transmitted.

The bit rate necessary to achieve target detection was computed and registered for each sequence of moving targets. The winner should be the coder producing the lower average bit rate for achieving target detection. The experiment was performed on a dataset composed of 12 standard $256 \times 256$ motion sequences of military targets on a natural background. Both camera motion and target motion are present in the sequences.

Table III shows the compression ratio and kbps, 10fps, $256 \times 256$ pixels/frame, in order to achieve target detection on each motion sequence. From this table we learn that the average bit rate to achieve target detection is 23.72 kbps using MC 3D-SPIHT and 12.60 kbps using REVIC.

Table III:  

Compression ratio and kbps, 10fps, $256 \times 256$ pixels/frame
at which targets are visible using REVIC/MC 3D-SPIHT
Motion Sequence	CODER
	REVIC		MC 3D-SPIHT
	Compr. ratio	kbps;10fps	Compr. ratio	kbps;10fps
1	512:1	10.24	402:1	13.04
2	492:1	10.66	132:1	39.72
3	512:1	10.24	252:1	20.80
4	252:1	20.80	202:1	25.95
5	412:1	12.73	412:1	12.73
6	512:1	10.24	302:1	17.36
7	492:1	10.66	152:1	34.49
8	512:1	10.24	72:1	72.82
9	202:1	25.95	202:1	25.95
10	212:1	24.73	152:1	34.49
11	372:1	14.09	172:1	30.48
12	512:1	10.24	202:1	25.95
Average	416:1	12.60	221:1	23.72

Figs. 1-9 show a visual comparison in performance of the MC 3D-SPIHT coder against the REVIC technique on motion sequences of military targets on a natural background. They illustrate frames from decoded motion sequences at the bit rate that is necessary to achieve target detection using the REVIC coder. They also display the respective frames using the MC 3D-SPIHT coder at the same bitrates. These figures demonstrate that the moving vehicles in each frame are easily visible using REVIC reconstructions at very low bit rates, whereas the moving targets are often invisible using MC 3D-SPIHT at the same bitrate. Recall that, in the REVIC codec, the embedded coding was performed without motion compensated filtering.

Figure 1: Frame 1 from decoded motion sequence at 10.24 kbps using MC 3D-SPIHT and REVIC, respectively.

Figure 2: Frame 1 from decoded motion sequence at 10.66 kbps using MC 3D-SPIHT and REVIC, respectively.

Figure 3: Frame 1 from decoded motion sequence at 10.24 kbps using MC 3D-SPIHT and REVIC, respectively.

Figure 4: Frame 5 from decoded motion sequence at 20.80 kbps using MC 3D-SPIHT and REVIC, respectively.

Figure 5: Frame 1 from decoded motion sequence at 12.73 kbps using MC 3D-SPIHT and REVIC, respectively.

Figure 6: Frame 1 from decoded motion sequence at 10.24 kbps using MC 3D-SPIHT and REVIC, respectively.

Figure 7: Frame 7 from decoded motion sequence at 10.66 kbps using MC 3D-SPIHT and REVIC, respectively.

Figure 8: Frame 1 from decoded motion sequence at 10.24 kbps using MC 3D-SPIHT and REVIC, respectively.

Figure 9: Frame 1 from decoded motion sequence at 14.09 kbps using MC 3D-SPIHT and REVIC, respectively.