Results of JSIVP Paper: "Feature-based detection and correction of occlusions and split of video objects"
  1. "Real" occlusions and splits:
    1. Occlusion is when one or more objects mask regions of other objects: Rate is 90% (18/20).
    2. Real splits due to deposit or object separation: Rate is 100% (6/6).
Sequence name Characteristics Result Occlusion detection rate Split detection rate
Total 90% (18/20) 100% (6/6)
Comm2 - Outdoor
- 604 frames
- Complete occlusion		 Comm2 100% (1/1) --
Street Survey - Outdoor
- Univ. of Rochester
- 1000 frames
- Multiple occlusions		 Street survey 100% (1/1) --
Urbi - Outdoor
- COST 211
- 300 frames
- Occlusions		 Urbi 100% (2/2) --
Road1 - Outdoor
- COST 211
- 300 frames
- Occlusions
- Split		 Road1 100% (1/1) 100% (1/1)
Hall-monitor - Indoor
- COST 211
- 300 frames
- Occlusions
- Split		 Hall-monitor 100% (1/1) 100% (1/1)
Meet-Split-3rdGuy - Indoor
- PETS 2004
- 929 frames
- Occlusions		 Meet split 3rd guy 67% (2/3) --
Left bag - Indoor
- PETS 2004
- 1444 frames
- Occlusion		 Left bag 100% (1/1) --
CU Hall - Indoor
- Concordia Univ.
- 690 frames
- Occlusion		 CU Hall 100% (4/4) --
Bishop entrance - Outdoor
- Concordia Univ.
- 1822 frames
- Occlusions
- Deposit 		Bishop entrance 75% (3/4) 100% (1/1)
INRS Hall - Indoor
- INRS
- 655 frames
- Deposit 		INRS Hall deposit -- 100% (1/1)
CUHall2 - Indoor
- Concordia Univ.
- 514 frames
- Occlusions
- Deposit 		CUHall deposit 100% (2/2) 100% (2/2)
  1. Occlusions and splits detected due to segmentation errors, e.g., objects enter the scene and are detected as multiple regions (i.e., fragmented)
Sequence name Result Details
Ekrlc EKRLC

Wrong occlusion detection: 1/0

The lady and her leg are detected as two separate objects and are tracked accordingly. When they merge later, this is detected as occlusion and handled as occlusion for some time before the system recovers.

Ekrlb EKRLC

Failed Split Detection: 1/0

Split due to segmentation errors at object 6 (green) where the foot splits and is too small to be detected as split

Ekrlb EKRLC

Successful split detection/correction

Testing the proposed algorithm with tough segmentation. Several splits were detected and corrected, e.g., the merging of regions in object 0 (red) with artificial straight lines in second 15, Note how the proposed algorithm provides reliable tracking despite flawed segmentation.