covdet − Covariant image frames detector
covdet [options] FILE ...
−v, −−verbose
Increase verbosity level (may be repeated).
−h, −−help
Show options and version.
−−output[=FILESPEC], −oFILESPEC
Enable/specify output file.
−−frames[=FILESPEC]
Enable/specify frames output.
−a,−−affine-adapt
Calc rotation covariant frames.
−r,−−orientation
Calc affine covariant frames.
−−method[DOG|HESSIAN], −m[DOG|HESSIAN]
Specify the method of frame localisation (’blobness’ measure).
−−descriptors[=FILESPEC]
Enable/specify descriptors output.
−−gss[=FILESPEC]
Enable/specify the Gaussian Scale Space (GSS) output. Exports each plane of the GSS as <input-image>_<octave-idx>_<level-idx>.pgm in binary format.
−−meta[=FILESPEC]
Enable/specify meta output.
−−first-octave=INTEGER
Specifiy the index of the first octave of the GSS.
−−edge-thresh=REAL
Specify the edge threshold.
−−peak-thresh=REAL
Specify the peak threshold.
−−magnif=REAL
Set the descriptor magnification factor.
−−norm-thresh=REAL
Set the minimum l2-norm of the descriptors before normalization. Descriptors below the threshold are set to zero.
−−calc-inv-sm
Export shape matrix of the ellipse frame as its inverse (Oxford format).
−−no-descriptors
Do not calculate descriptors.
−−win-size=REAL
Set the variance of the Gaussian window that determines the descriptor support. It is expressend in units of spatial bins. Used only for isotrpic frames (discs).
−−aff-win-size=REAL
Size of the window used for Second Moment Matrix (SMM) calculation. SMM is used as an estimation of the affine shape. Used only for anisotropic frames (ellipses).
−−aff-max-iter=REAL
Maximum number of iterations for the affine shape estimation. If the convergence criterion is not fulfilled in these steps, keypoint is rejected. Used only for anisotropic frames (ellipses).
−−aff-conv-thr=REAL
Convergence criterion for affine shape estimation. Iteration converges when the isotropy measure value, defined as ratio between second moment matrix eigen values, differs from the ideal isotropy measure by value smaller than this threshold.
−−sift-patch-size=REAL
Size of the patch used for SIFT descriptor calculation. Used only for anisotropic frames (ellipses).
covdet computes the Covariant image frames. In the simplest case, sift reads an image file (in PGM format), computes the oriented-disc frames using difference of gaussian image response (in fact frames produced by SIFT detector) and their descriptors and write them to a file. Alternatively, --frames option can be used to suppress the standard output and produce a file with the feature frames only. Similarly, --descriptors suppresses the standard output and produce a file of descriptors only. --output can be used in combination with the previous too if the combined output of frames and descriptors is required.
In the default settings the input image is upscaled. To suppress this run with parameter --first-octave=0
The response function used for detection of distinct image regions can be defined using --resp-func parameter. The allowed values are DOG and HESSIAN computing Difference of Gaussian (SIFT response function) or determinant of Hessian matrix respectively.
Also the type of detected frame can be defined using --orientation and --affine-adapt parameters where frames of type Disc (no orient., no affine), Oriented disc (rot. covariant, no affine), Ellipse, (no orient., affine covariant), Oriented ellipse, (rot. covariant, affine covariant),
A Disc frame is specified by 3 floating point numbers (in order: coordinate of the center x and y, and Sigma, radius of the region in the image).
An Oriented disc frame is specified by 4 floating point numbers (in order: coordinate of the center x and y, and elements Sigma, radius of the region in the image and orientation Theta in the image).
An Ellipse frame is specified by 5 floating point numbers (in order: coordinate of the center x and y, and elements E11, E21, E22 of the shape matrix which is inverse of Ellipse matrix). The shape matrix can be calculated as an Ellipse matrix (not its inverse) when parameter --calc-inv-sm is defined.
An Oriented ellipse frame is specified by 6 floating point numbers (in order: coordinate of the center x and y, and elements A11, A21, A21, A22 of the matrix mapping the unit circle to the ellipse).
A feature descriptor is an histogram. It is specifyied by 128 non-negative integer numbers.
covdet can
process multiple images and derive the names of the various output
files by means of FILESPEC (see vlfeat(7)). Both
frames and descriptors can be saved/loaded either in ascii or
binary format.
Ascii format
Each descriptor is a list of 128 decimal numbers and occupies a line of text. Each frame is a list of 5 numbers, specifying an elliptical frame, one per line of text. The combined output file (with both frames and descriptors) stores a frame and the corresponding descriptor for each line of text.
Binary format
The binary format is similar to the ASCII format, except that each frame element is stored as an IEEE double floating point number (eight bytes) and each descriptor element is stored as an unsiged integer (one byte). The data is written in little endian order.
covdet test.pgm --peak-thresh=0.03
Caluclates the SIFT frames and descriptors of image test.pgm with peak threshold equal to 0.03 and save them back to test.sift in ascii format.
covdet test.pgm --no-descriptors
Writes the frames to test.sift in ascii format without calculating descriptors.
covdet test.pgm --descriptors=bin://
writes only the frames to test.descr. in binary format.
covdet -r hessian -f ellipse --calc-inv-sm test.pgm
detect hessian-affine frames and descriptors and stores them to test.sift. file. The shape matrix is exported as its inverse (Oxford format - exported as the ellipse matrix).
covdet -f dics test.pgm −−frames=my.frames -o bin:///tmp/%.sift --gss
Computed upright sift frames and descriptors writes the frames in ascii format to my.frames and the combined frames-descriptors file in binary format to /tmp/test.sift. Exports the GSS planes into set of images test_[-1..num-octaves]_[-01..03].pgm Where the default number of levels is three.
vlfeat(7).
D. G. Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2(60):91-110, 2004.
Mikolajczyk, K. and Schmid, C. Scale & Affine Invariant Interest Point Detectors IJCV, (60), 2004.