search

lifeng9472 / STRCF

Learning Spatial-Temporal Regularized Correlation Filters for Visual Tracking (CVPR 2018)
149 stars 40 forks source link

Interpolation Function #7

Open iou1992 opened 6 years ago

iou1992 commented 6 years ago

Hi, After add interpolation function in the C-COT when the the gaussian label function donot change ,why the performance will reduce ? thx advance

lifeng9472 commented 6 years ago

You mean replacing the 'resp_newton' function with the 'optimize_scores' used in CCOT? That surely might bring some changes due to the different processing of response scores. But I think it shouldn't make much difference on the whole OTB-100 dataset.

iou1992 commented 6 years ago

No. % Compute Fourier series of interpolation function [interp1_fs, interp2_fs] = cellfun(@(sz) get_interp_fourier(sz, params), filter_sz_cell, 'uniformoutput', false); .... % Interpolate features to the continuous domain xtf = interpolate_dft(xtf, interp1_fs, interp2_fs);

the code is CCOT tracker 。 i think the fuction of interpolate_dft is useful for STRCT . But the performence is becoming bad if add the function of interpolate_dft . thanks.

lifeng9472 commented 6 years ago

Actually, from my previous experiments on combing CCOT and STRCF methods, the introduction of continuous convolution won't bring much improvements for STRCF. As for your question, maybe you should also interploate the label y as CCOT had done. Futhermore, the hyper-parameters also needs to be tuned for incorporating the continous convolution.

iou1992 commented 6 years ago

thank you for your advice.