DSPSIFT detects less features that SIFT #1777
Comments
|
Interesting; would certainly be worth investigating. |
|
Anything I can help on this, I can provide some comparisons for the number of features extracted with different settings. But how that affect the final reconstruction is hard to measure, when a dataset is well surveyed, 4000 or 8000 features might not have big difference with respect to accuracy. |
|
Comparison numbers would be interesting. Attaching the PDF reports would also provide some metrics on the feature quality between comparisons. |
|
The tests are performed on Brighton beach with default settings and set min-num-features to 50000 so it can exhaustively find features at peak threshold 0.0001. I'll use dspsift(0) to indicate first octave 0 and dspsift(-1) for first octave -1.
The table can show that the dspsift(0) has basically one quality level less features comparing to sift and dspsift(-1). But dspsift(-1) will definitely cost more processing time. Some logs for processing time below DSPSIFT(first-octave:-1, feature-quality:high, min-num-features:50000) DSPSIFT(first-octave:0, feature-quality:high, min-num-features:50000) report SIFT(feature-quality:high, min-num-features:10000) report DSPSIFT(first-octave:-1, feature-quality:high, min-num-features:10000) report DSPSIFT(first-octave:0, feature-quality:high, min-num-features:10000) |
|
Nice, thanks for the detailed data reports! Almost all metrics in the report for dspsift(-1) indicate an improvement in matching. An increase in runtime is expected, but seems reasonable especially when compared to sift. I think this can be patched in our OpenSfM fork and brought into ODM. 👍 If you can open a PR there with the proposed change targeting the |
|
OK, I'll create a PR for this. I also noticed that the max number of features is 1.5 times of min-num-features in sift, do we need to have the same behavior in dspsift? |
|
I don't think that's required. |
|
#1780 closes this. Thanks for the research and fix @originlake ! |
Hi,
The implementation of dspsift use first octave 0 and I believe OpenCV's first octave is -1. This makes the difference that the medium quality of SIFT is kinda on par with high quality in dspsift (due to the upscaling of -1 first octave). I can observe the difference during processing that the SIFT usually detects more features than dspsift, and when I set the first octave of dspsift to -1, I can then get similar amount of features. Would it be better to make them comparable at the same quality settings?
(Opencv's default firstOctave)
https://github.com/opencv/opencv/blob/c6ace77e21d9db3fcee5953ef36d5ba7ec289995/modules/features2d/src/sift.dispatch.cpp#L508
(DSPSIFT implementation)
https://github.com/OpenDroneMap/OpenSfM/blob/792b38714e42588b6f0e9354527660c7e54f18b9/opensfm/src/features/src/dspsift.cc#L76
The text was updated successfully, but these errors were encountered: