HI, have a dataset of 6 views at different positions.
I ran IncrementalSFM it showed no error, but only 4 camera positions were estimated. In the camera_poses.json file there are extrinsics for 4 views (2 are missing). Is it normal?
When I ran GlobalSFM, I got the error: Rotation Averaging failure. I tried with the options "-r 1" and "-r 2", both had the same problem.
Any suggestions would be appreciated!
squashking
All 22 comments
Please try ANN_L2 matcher.
whuaegeanse
on 15 Jan 2018
@whuaegeanse thanks. I tried add "-n ANNL2" when running openMVG_main_ComputeMatches . But still got the same error.
squashking
on 15 Jan 2018
Show more information, eg log ,and share your data.
whuaegeanse
on 15 Jan 2018
@whuaegeanse Thanks a lot! Below is the log when I run the commands (I don't know how to share the data here, please let me know if i have to share data as well):
You called :
openMVG_main_SfMInit_ImageListing
--imageDirectory /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/images/
--sensorWidthDatabase /home/turtlebot/openMVG/src/openMVG/exif/sensor_width_database/sensor_width_camera_database.txt
--outputDirectory /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches
--focal 1400
--intrinsics
--camera_model 3
--group_camera_model 0
- Image listing -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
SfMInit_ImageListing report:
listed #File(s): 6
usable #File(s) listed in sfm_data: 6
usable #Intrinsic(s) listed in sfm_data: 6
You called :
openMVG_main_ComputeFeatures
--input_file /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches/sfm_data.json
--outdir /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches/
--describerMethod SIFT
--upright 0
--describerPreset HIGH
--force 0
--numThreads 0
- EXTRACT FEATURES -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
Task done in (s): 0
You called :
openMVG_main_ComputeMatches
--input_file /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches/sfm_data.json
--out_dir /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches/
Optional parameters:
--force 0
--ratio 0.8
--geometric_model e
--video_mode_matching -1
--pair_list
--nearest_matching_method ANNL2
--guided_matching 0
--cache_size unlimited
- Regions Loading -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
PUTATIVE MATCHES -
PREVIOUS RESULTS LOADED; #pair: 15- Geometric filtering -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
- Geometric filtering -
Task done in (s): 0
Export Adjacency Matrix of the pairwise's geometric matches
Global Structure from Motion:
Open Source implementation of the paper:
"Global Fusion of Relative Motions for Robust, Accurate and Scalable Structure from Motion."
Pierre Moulon, Pascal Monasse and Renaud Marlet. ICCV 2013.
- Features Loading -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
CleanGraph_KeepLargestBiEdge_Nodes():: => connected Component: 6
Connected component of size: 1
Connected component of size: 1
Connected component of size: 1
Connected component of size: 1
Connected component of size: 1
Connected component of size: 1
- Relative pose computation -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
Relative motion computation took: 0(ms)
GlobalSfM:: Rotation Averaging failure!
squashking
on 16 Jan 2018
In the Global SFM step, the connected component is 6 and each component only has 1 image. In this case Gloabal SFM cannot be used for that Global SFM need 3 or more images in a connected component.
Please try ULTRA mode in feature extraction step and ANN_L2 matcher in feature matching step.
whuaegeanse
on 16 Jan 2018
@whuaegeanse I tried that but still got the error. Do you mean that 6 images are not enough to do global SFM? Here is the log:
You called :
openMVG_main_SfMInit_ImageListing
--imageDirectory /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/images/
--sensorWidthDatabase /home/turtlebot/openMVG/src/openMVG/exif/sensor_width_database/sensor_width_camera_database.txt
--outputDirectory /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches
--focal 1400
--intrinsics
--camera_model 3
--group_camera_model 0
- Image listing -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
SfMInit_ImageListing report:
listed #File(s): 6
usable #File(s) listed in sfm_data: 6
usable #Intrinsic(s) listed in sfm_data: 6
You called :
openMVG_main_ComputeFeatures
--input_file /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches/sfm_data.json
--outdir /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches/
--describerMethod SIFT
--upright 0
--describerPreset ULTRA
--force 0
--numThreads 0
- EXTRACT FEATURES -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
Task done in (s): 0
You called :
openMVG_main_ComputeMatches
--input_file /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches/sfm_data.json
--out_dir /home/turtlebot/openMVG_Build/software/SfM/Test_data8asp_nolight/matches/
Optional parameters:
--force 0
--ratio 0.8
--geometric_model e
--video_mode_matching -1
--pair_list
--nearest_matching_method ANNL2
--guided_matching 0
--cache_size unlimited
- Regions Loading -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
PUTATIVE MATCHES -
PREVIOUS RESULTS LOADED; #pair: 15- Geometric filtering -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
- Geometric filtering -
Task done in (s): 0
Export Adjacency Matrix of the pairwise's geometric matches
Global Structure from Motion:
Open Source implementation of the paper:
"Global Fusion of Relative Motions for Robust, Accurate and Scalable Structure from Motion."
Pierre Moulon, Pascal Monasse and Renaud Marlet. ICCV 2013.
- Features Loading -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
CleanGraph_KeepLargestBiEdge_Nodes():: => connected Component: 6
Connected component of size: 1
Connected component of size: 1
Connected component of size: 1
Connected component of size: 1
Connected component of size: 1
Connected component of size: 1
- Relative pose computation -
0% 10 20 30 40 50 60 70 80 90 100%
|----|----|----|----|----|----|----|----|----|----|
Relative motion computation took: 0(ms)
GlobalSfM:: Rotation Averaging failure!
squashking
on 16 Jan 2018
It seems in your case that no matches are found at the geometric matching step so perhaps there is not a lot of features matches even at the putative matches step.
Can you try with the sequential pipeline with the preset HIGH or ULTRA?
Feel free to share your image dataset and we can take a look to it.
Tip for photogrammetry there is some important points:
- the better the intrinsic value are, the better the 3d point cloud and camera location will be
- image with high overlap and texture area are often required in order to connect the pictures together.
pmoulon
on 16 Jan 2018
@pmoulon Yes I used ULTRA in the above run. Could you tell me how to share the image data? Thanks a lot!
squashking
on 16 Jan 2018
You can make an archive (zip) of your images and drag and drop in your GitHub message or you can send your image to the private mailing list openmvg-team[AT]googlegroups.com
pmoulon
on 16 Jan 2018
@pmoulon Hi, I've sent the images to the googgroups email since it's not allowed to send files larger than 10MB here. There are only 6 images taken by one ZED camera. The image files do not have EXIF information. The pixel dimension is 0.002mm*0.002mm, the focal lenght in pixel is 1400, so I used the "-f 1400" when running Imagelisting. Thank you very much!
squashking
on 16 Jan 2018
@squashking Before you redo the pipeline , please remove the match directory . Or use --force 1 in the feature extraction step.
whuaegeanse
on 16 Jan 2018
@whuaegeanse I just tried that and got the same error.
squashking
on 16 Jan 2018
Share your data to me. [email protected]
whuaegeanse
on 16 Jan 2018
Thanks @squashking I will have a look to your dataset.
pmoulon
on 16 Jan 2018
@whuaegeanse I have sent you the dataset. Thank you very much!
squashking
on 16 Jan 2018
Thanks @squashking .
whuaegeanse
on 17 Jan 2018
Since your baseline is large it is better to use the Sequential pipeline for this dataset.
Here the results with the default setting (-f 1400 and default sequentialSfM)
4 camera registered out of 6
Here the results with (-f 1400, -m AKAZE_FLOAT, -p HIGH, -n ANNL2 and sequential SfM)
6 camera registered out of 6
colorized.ply.zip
So as conclusion, if you want to use the Global SfM module you need a smaller baseline when you are moving. Note that here we don't have dense RGBD data for registration of continuous scan, we only rely on sparse feature correspondences.
pmoulon
on 17 Jan 2018
@pmoulon Thank you so much! But I've got a newbie question. What is baseline? Do you mean the distance between neighbouring views? And what is the sequential SFM pipline? Do you mean IncrementalSFM? Thanks again!
squashking
on 18 Jan 2018
Baseline was to express the distance between two camera.
SequentialSfM and Incremental SfM are the same (depict the fact that do thing incrementally or sequentially).
pmoulon
on 18 Jan 2018
@pmoulon May I know how did you get he colorized point cloud? Did you use some dense reconstruction tool? thanks.
squashking
on 18 Jan 2018
@squashking openMVG contain a function to do that "openMVG_main_ComputeSfM_DataColor"
MesHeritage
on 21 Jan 2018
Closing the issue since now we now that the global SfM engine is failing due to lack of feature matches between some images (The global SfM engine use triplet of images).
pmoulon
on 22 Jan 2018
Related issues
lab3d
·
7Comments
yuancaimaiyi
·
4Comments
yuyou
·
6Comments
CanCanZeng
·
3Comments
snowzurfer
·
3Comments
Most helpful comment
Since your baseline is large it is better to use the Sequential pipeline for this dataset.
Here the results with the default setting (-f 1400 and default sequentialSfM)
4 camera registered out of 6
Here the results with (-f 1400, -m AKAZE_FLOAT, -p HIGH, -n ANNL2 and sequential SfM)
6 camera registered out of 6
colorized.ply.zip
So as conclusion, if you want to use the Global SfM module you need a smaller baseline when you are moving. Note that here we don't have dense RGBD data for registration of continuous scan, we only rely on sparse feature correspondences.