This webpage contains additional material (videos) for the article:

G.D. Evangelidis and C. Bauckhage, "Efficient Subframe Video Alignment using Short Descriptors", IEEE Transaction on Pattern Analysis and Machine Intelligence, vol.35, no. 10, Oct. 2013, pages 2371-2386 (pdf)

Video results for:
Moving Cameras
Backroad Sequence (moving cameras)
Campus Sequence (moving cameras)
Highway Sequence (moving cameras)
Suburb Sequence (moving cameras, appearance variation)
Pedzone2 Sequence (moving cameras, appearance variation)
Static cameras (non-rigid scene motion)
Wind Sequence (static cameras, non-rigid scene motion)
Water Sequence (static cameras, non-rigid scene motion)
Inria Sequence (static cameras, non-rigid scene motion, appearance variation)

Moving Cameras

Backroad Sequence

This text will be replaced			This text will be replaced
Reference Sequence			Input sequence

Synchronization:

In these representations, the top frame is the input(query) frame while the bottom frame is the synchronized reference image obtained by each method.

This text will be replaced

Spatio-temporal alignment:

Fusion videos have been created by combining the R and B channels from the input sequence and G component from the reference sequence, but warped in space and time w.r.t the outcome of each algorithm. This way changes are identified with lawn-green and hot-pink colors.

This text will be replaced

Proposed Quad-Tree

SIFT-BoW + SIFT-flow [1],[3]

Map-Inference + Lucas-Kanade[2]

Caspi-Irani[4]

Campus Sequence

This text will be replaced			This text will be replaced
Reference Sequence			Input sequence

Synchronization:

In these representations, the top frame is the input(query) frame while the bottom frame is the synchronized reference image obtained by each method.

This text will be replaced

Spatio-temporal alignment:

This text will be replaced

Proposed Quad-Tree

SIFT-BoW + SIFT-flow [1],[3]

Map-Inference + Lucas-Kanade[2]

Caspi-Irani[4]

Highway Sequence

This text will be replaced			This text will be replaced
Reference Sequence			Input sequence

Synchronization:

In these representations, the top frame is the input(query) frame while the bottom frame is the synchronized reference image obtained by each method.

This text will be replaced

Spatio-temporal alignment:

This text will be replaced

Proposed Quad-Tree

SIFT-BoW + SIFT-flow [1],[3]

Map-Inference + Lucas-Kanade[2]

Caspi-Irani[4]

Pedzone-2 Sequence (with appearance variation)

This text will be replaced			This text will be replaced
Reference Sequence			Input sequence

Spatio-temporal alignment:

This text will be replaced				This text will be replaced				This text will be replaced
Proposed Quad-Tree				SIFT-BoW + SIFT-flow [1],[3]				Caspi-Irani[4]

Suburb Sequence (different weather conditions)

This text will be replaced			This text will be replaced
Reference Sequence			Input sequence

Spatio-temporal alignment:

This text will be replaced				This text will be replaced				This text will be replaced
Proposed Quad-Tree				SIFT-BoW + SIFT-flow [1],[3]				Caspi-Irani[4]

Non-rigid scene motion - Static cameras

In this comparison, sequences have been synchronized by the proposed Quad-Tree and a RANSAC-based line fitting, hence they are synchronized up to frame accuracy. Then, starting from the same temporal intialization, we test the proposed spatio-temporal ECC scheme and the algorithm proposed by Caspi-Irani. Note that when the algorithms use 1 frame, it means that a frame-to-subframe alignment scheme is established (it is spatio-temporal as well). We provide the uncompressed videos for the cases that the difference in performance may not be obvious.

Wind sequence

Video 1			Video2
video_wind_1			video_wind_2

Spatio-temporal alignment using subsequences

Proposed 1 frame	Proposed 3 frames	Proposed 9 frames
video_wind_proposed_1	video_wind_proposed_3	video_wind_proposed_9
Caspi-Irani 1 frame	Caspi-Irani 3 frames	Caspi-Irani 9 frames
video_wind_caspi_1	video_wind_caspi_3	video_wind_caspi_9

Uncompressed video files:
Proposed (9 frames)
Caspi-Irani (9 frames)

Water sequence

Video 1			Video2
video_water_1			video_water_2

Spatio-temporal alignment using subsequences

Proposed 1 frame	Proposed 3 frames	Proposed 9 frames
video_water_proposed_1	video_water_proposed_3	video_water_proposed_9
Caspi-Irani 1 frame	Caspi-Irani 3 frames	Caspi-Irani 9 frames
video_water_caspi_1	video_water_caspi_3	video_water_caspi_9

Uncompressed video files:
Proposed (9 frames) (70 MB)
Caspi-Irani (9 frames) (70 MB)

Uncompressed video files:
Proposed (3 frames) (70 MB)
Caspi-Irani (3 frames) (70 MB)

Inria sequence (with appearance variation)

Video 1			Video2
video_inria_1			video_inria_2

Spatio-temporal alignment using subsequences

Proposed 1 frame	Proposed 3 frames	Proposed 9 frames
video_inria_proposed_1	video_inria_proposed_3	video_inria_proposed_9
Caspi-Irani 1 frame	Caspi-Irani 3 frames	Caspi-Irani 9 frames
video_inria_caspi_1	video_inria_caspi_3	video_inria_caspi_9

Uncompressed video (avi) files:
Proposed (9 frames) (60 MB)
Caspi-Irani (9 frames) (60 MB)

[1] J. Sivic and A. Zissermann, Efficient Visual Search of videos cast as text retrieval", IEEE Trans. on Pattern Analysis & Machine Itelligence, vol.31, no. 4, pp. 591-606, 2009
[2] F. Diego, D. Ponsa, J. Serrat and A. Lo'pez, Video Alignment for change detecion, IEEE Trans. on Image Processing, vol. Preprint, Issue 99, 2010
[3] C. Liu, J. Yuen, A. Torralba, J. Sivic, W. T. Freeman, SIFT flow: dense correspondence across difference scenes, in ECCV 2008
[4] Y. Caspi, M. Irani, Spatio-temporal alignment of sequences, IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 24, no. 11, pp. 1409-1424, 2002