vot 77503 preparation and characterization of cation exchange ...
The Visual Object Tracking VOT-TIR2015 Challenge...
Transcript of The Visual Object Tracking VOT-TIR2015 Challenge...
![Page 1: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/1.jpg)
The Visual Object Tracking VOT-TIR2015 Challenge ResultsMichael Felsberg, Amanda Berg, Jörgen Ahlberg, Gustav Häger, Matej Kristan, Jiři Matas, AlešLeonardis, Luka Čehovin, Gustavo Fernández, TomášVojiř, Georg Nebehay, Roman Pflugfelder, et al.
![Page 2: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/2.jpg)
Outline
1. Scope of the VOT-TIR challenge
– Thermal infrared imaging
2. VOT-TIR2015 challenge overview
– Evaluation system
– Dataset
– Performance evaluation measures
3. VOT-TIR2015 results overview
4. Summary and outlook
Felsberg et al., VOT-TIR2015 results 2
![Page 3: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/3.jpg)
Scope of the VOT-TIR challenge
• Single-object, single thermal infrared (TIR) camera, model-free, short-term, causal trackers
• Model-free:
– Nothing but a single training example is provided by the BBox in the first frame
• Short-term:
– Tracker does not perform re-detection
– Once it drifts off the target we consider that a failure
• Causality:
– Tracker does not use any future frames for pose estimation
• Object state defined as an upright bounding box
Felsberg et al., VOT-TIR2015 results 3
![Page 4: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/4.jpg)
Felsberg et al., VOT-TIR2015 results 4
Thermal Infrared
htt
p:/
/ww
w.t
hes
es.u
lava
l.ca/
20
05
/23
01
6/a
pb
.htm
l
![Page 5: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/5.jpg)
Applications of TIR
• Scientific research
• Security
• Fire monitoring
• Search and rescue
• Automotive safety
• Personal use
• Military
Felsberg et al., VOT-TIR2015 results 5
![Page 6: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/6.jpg)
Why a separate challenge?
Tracking in TIR different from tracking in lowresolution grayscale visual?
Many similarities but also interesting differences
• 16-bit
• Constant values if radiometric
• Less structure/edges/texture
• No shadows
• Noise: blooming, resolution, dead pixels
Felsberg et al., VOT-TIR2015 results 6
![Page 7: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/7.jpg)
Evaluation system from VOT 2015
• Matlab-based kit to automatically perform a battery of standard experiments
• Download from our homepage
– https://github.com/votchallenge/vot-toolkit
– select the vottir2015 experiment stack
• Plug and play!
– Supports multiple platforms and programming languages (C/C++/Matlab/Python, etc.)
• Easy to evaluate your tracker on our benchmarks
• Deep integration with tracker - Fast execution of experiments
Felsberg et al., VOT-TIR2015 results 7
![Page 8: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/8.jpg)
Felsberg et al., VOT-TIR2015 results 8
VOT-TIR2015 Dataset: LTIR
• Follows VOT approach:
– Keep it sufficiently small, diverse and well annotated
– Follow the VOT dataset construction methodology
• Linköping Thermal InfraRed (LTIR) datasetA. Berg, J. Ahlberg, M. Felsberg, A Thermal Object Tracking Benchmark. AVSS 2015.
![Page 9: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/9.jpg)
Existing datasets
Felsberg et al., VOT-TIR2015 results 9
LITIVOSU Pedestrian OSU Color-Thermal
BU-TIV
![Page 10: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/10.jpg)
Felsberg et al., VOT-TIR2015 results 10
• Different sources
• Different applications
• Different sensors
• Moving + stationary sensors
• Radiometric + non-radiometric
• 8/16 bits
![Page 11: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/11.jpg)
Felsberg et al., VOT-TIR2015 results 11
• Different sources
• Different applications
• Different sensors
• Moving + stationary sensors
• Radiometric + non-radiometric
• 8/16 bits
![Page 12: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/12.jpg)
Felsberg et al., VOT-TIR2015 results 12
• Different sources
• Different applications
• Different sensors
• Moving + stationary sensors
• Radiometric + non-radiometric
• 8/16 bits
![Page 13: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/13.jpg)
Felsberg et al., VOT-TIR2015 results 13
• Different sources
• Different applications
• Different sensors
• Moving + stationary sensors
• Radiometric + non-radiometric
• 8/16 bits
![Page 14: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/14.jpg)
Properties
• 20 Sequences
• Average sequence length 563
• Annotations in accordance with VOT-standard
– Bounding-box
– 11 global attributes (per-sequence)
– 6 local attributes (per-frame)
Felsberg et al., VOT-TIR2015 results 14
Occlusion, dynamics change, object motion, object size change, camera motion, neutral
![Page 15: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/15.jpg)
Sequence details
Felsberg et al., VOT-TIR2015 results 15
ASL-TID
![Page 16: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/16.jpg)
Felsberg et al., VOT-TIR2015 results 16
![Page 17: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/17.jpg)
Felsberg et al., VOT-TIR2015 results 17
OSU Pedestrian
Ours Old
![Page 18: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/18.jpg)
Global attributes
Felsberg et al., VOT-TIR2015 results 18
![Page 19: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/19.jpg)
LTIR general stats
Felsberg et al., VOT-TIR2015 results 19
![Page 20: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/20.jpg)
Felsberg et al., VOT-TIR2015 results 20
![Page 21: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/21.jpg)
Will it be different? Test against VOT2014
Felsberg et al., VOT-TIR2015 results 21
VOT2014 LTIR
![Page 22: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/22.jpg)
Performance evaluation measures
• Basically the same as VOT2015 (based on 8-bit)
– accuracy
– robustness
• Evaluated globally and per-attribute
– raw value
– rank
• Overall: expected average overlap
• Speed in EFO units
Felsberg et al., VOT-TIR2015 results 22
![Page 23: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/23.jpg)
Results
• 20 submitted trackers
• 4 added by VOT committee
• 20 of 24 trackers in both challenges
• Various classes of trackers
– mean shift extensions (ASMS, PKLTF, SumShift, DTracker)
– part-based trackers (LDP, G2T, AOGTracker, MCCT, FoT)
– correlation filter based (NSAMF, OACF, SRDCFir, sKCF, STC, MKCF+, CCFP, SME, KCFv2)
– others (EBT, CMIL, sPST, Struck, ABCD, HotSpot)
Felsberg et al., VOT-TIR2015 results 23
![Page 24: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/24.jpg)
Results (sequence pooling)
Felsberg et al., VOT-TIR2015 results 24
![Page 25: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/25.jpg)
Results (attribute normalization)
Felsberg et al., VOT-TIR2015 results 25
![Page 26: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/26.jpg)
Felsberg et al., VOT-TIR2015 results 26
ECCV’14VOT’14(*)
(*) improved version of SAMF
![Page 27: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/27.jpg)
Felsberg et al., VOT-TIR2015 results 27
![Page 28: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/28.jpg)
Felsberg et al., VOT-TIR2015 results 28
![Page 29: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/29.jpg)
Sequence ranking
• A_f: average number of trackers failed per frame
• M_f: max. number of trackers failed at a single frame
Felsberg et al., VOT-TIR2015 results 29
Sequence Scorecrowd 2quadrocopter 2,5quadrocopter2 2,5garden 3mixed_distractors 3saturated 3,5selma 3,5street 3,5birds 4crouching 4
Sequence Scorejacket 4hiding 4,5car 5crossing 5depthwise_crossing 5horse 5rhino_behind_tree 5running_rhino 5soccer 5trees 5
challenging:0.06<=A_f<=0.214<=M_f<=22
intermediate:0.04<=A_f<=0.1
6<=M_f<=11easiest:
0<=A_f<=0.040<=M_f<=7
![Page 30: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/30.jpg)
Difficulty analysis
Felsberg et al., VOT-TIR2015 results 30
![Page 31: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/31.jpg)
Summary
• New challenge with LTIR dataset
• Dataset too easy (or trackers too good)
• Large fraction of trackers show similar ranking in both challenges
– in contrast to VOT2014
• Top-perforing triple: SRDCFir, sPST, MCCT
• Best real-time method: sKCF
• Available at http://www.votchallenge.net/vot2015
Felsberg et al., VOT-TIR2015 results 31
![Page 32: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/32.jpg)
Winners of the VOT-TIR2015 Challenge:
Yang Hua, Karteek Alahari, and Cordelia Schmid:
Simplified Proposal SelectionTracker (sPST)
Presentation at VOT2015 today at 17:35
Award sponsored by
Felsberg et al., VOT-TIR2015 results 32
![Page 33: The Visual Object Tracking VOT-TIR2015 Challenge Resultsdata.votchallenge.net/vot2015/presentations/vot_tir_2015_presentati… · The Visual Object Tracking VOT-TIR2015 Challenge](https://reader033.fdocuments.in/reader033/viewer/2022053007/5f0b290f7e708231d42f2550/html5/thumbnails/33.jpg)
Thanks
• The VOT2015 Committee
• Jörgen and Amanda
• Everyone who participated:
Felsberg et al., VOT-TIR2015 results 33
Alan Lukezic, Alvaro Garcia-Martin, Amir Saffari, Ang Li, Andres Solis Montero,Baojun Zhao, Cordelia Schmid, Dapeng Chen, Dawei Du, Fahad Shahbaz Khan, FatihPorikli, Gao Zhu, Guibo Zhu, Hanqing Lu, Hilke Kieritz, Hongdong Li, HonggangQi, Jae-chan Jeong, Jae-il Cho, Jae-Yeong Lee, Jianke Zhu, Jiatong Li, Jiayi Feng,Jinqiao Wang, Ji-Wan Kim, Jochen Lang, Jose M. Martinez, Kai Xue, Karteek Alahari, LiangMa, Lipeng Ke, Longyin Wen, Luca Bertinetto, Martin Danelljan, Michael Arens, MingTang, Ming-Ching Chang, Ondrej Miksik, Philip H S Torr, Rafael Martin-Nieto, RobertLaganiere, Sam Hare, Siwei Lyu, Song-Chun Zhu, Stefan Becker, Stephen L Hicks, StuartGolodetz, Sunglok Choi, Tianfu Wu, Wolfgang Hubner, Xu Zhao, Yang Hua, Yang Li,Yang Lu, Yuezun Li, Zejian Yuan, and Zhibin Hong