ECO跑UAV数据集填坑

 

首先有读懂以下解决方案的基础：

https://blog.csdn.net/qq_33970986/article/details/105496382

按照这种方式修改之后，仍然出现error

遇到这种情况，主要是因为tracker中最后一行[seq, results] = get_sequence_results(seq)的问题，打开这个函数，

function [seq, results] = get_sequence_results(seq)
% »ñÈ¡ÐòÁнá¹û
if strcmpi(seq.format, 'otb')%otbÊý¾Ý¼¯
    results.type = 'rect';
    results.res = seq.rect_position;
elseif strcmpi(seq.format, 'vot')%votÊý¾Ý¼¯
    seq.handle.quit(seq.handle);
else
    error('Uknown sequence format');%δ֪Êý¾Ý¼¯
end
 
if isfield(seq, 'time') %È·¶¨ÊäÈëÊÇ·ñΪ½á¹¹Êý×é×ֶΣ¬Êä³öÂß¼­0-1
    results.fps = seq.num_frames / seq.time;%ÇóÖ¡ÂÊ
else
    results.fps = NaN;
end
 
%AUC results
pd_boxes = results.res;
thresholdSetOverlap = 0: 0.05 : 1;
success_num_overlap = zeros(1, numel(thresholdSetOverlap));
res = calcRectInt(seq.gt_box, pd_boxes);
for t = 1: length(thresholdSetOverlap)
    success_num_overlap(1, t) = sum(res > thresholdSetOverlap(t));
end
results.cur_AUC = mean(success_num_overlap) / size(seq.gt_box, 1);

 

会发现它没有对UAV数据及进行定义，并且提前计算了AUC，

通过main_unning的错误报告问题可知seq.gt_box未定义：

对照错误信息与BACF的运行代码进行修补，

在BACF中，最后的结果是以这种形式体现的：

因此，需要将ECO中的[seq, results] = get_sequence_results(seq)

修改一下：

1，加入uav选项，其实跟otb一样都是rect

2，定义好结果，不需要函数内自行比对AUC，删除这一块

3，fps计算方式无误，不需要修改

function [seq, results] = get_sequence_results(seq)
% »ñÈ¡ÐòÁнá¹û
if strcmpi(seq.format, 'otb')%otbÊý¾Ý¼¯
    results.type = 'rect';
    results.res = seq.rect_position;
elseif strcmpi(seq.format, 'vot')%votÊý¾Ý¼¯
    seq.handle.quit(seq.handle);
elseif strcmpi(seq.format, 'uav')%uavÊý¾Ý¼¯
    results.type = 'rect';
    results.res = seq.rect_position;
else
    error('Uknown sequence format');%δ֪Êý¾Ý¼¯
end
 
if isfield(seq, 'time') %È·¶¨ÊäÈëÊÇ·ñΪ½á¹¹Êý×é×ֶΣ¬Êä³öÂß¼­0-1
    results.fps = seq.num_frames / seq.time;%ÇóÖ¡ÂÊ
else
    results.fps = NaN;
end
 
% %AUC results
% pd_boxes = results.res;
% thresholdSetOverlap = 0: 0.05 : 1;
% success_num_overlap = zeros(1, numel(thresholdSetOverlap));
% res = calcRectInt(seq.gt_box, pd_boxes);
% for t = 1: length(thresholdSetOverlap)
%     success_num_overlap(1, t) = sum(res > thresholdSetOverlap(t));
% end
% results.cur_AUC = mean(success_num_overlap) / size(seq.gt_box, 1);

完成，重新跑一下试试：

完美运行，所以其实就是get_sequence_results中多此一举，多了一个计算AUC的步骤，因为读取不到groundtruth导致函数无法返回gt_box进而引起评测代码不能正确读取结果而导致的报错。