visdrone / visdrone2018-mot-toolkit Goto Github PK

@VisDrone When I eval task4a，and some error happened：

evaluating VisDrone2019 Task4 :: loading groundtruth
evaluating VisDrone2019 Task4 :: loading predictions
错误使用 evaluateTrackA (第 130 行)
Find multiple labels in a tracklet.

出错 evalMOT (第 31 行)
[ap, recall, precision] = evaluateTrackA(seqPath, resPath, gtPath, evalClassSet);

Could you help me ?

Hello

I am trying to find out how many different classes of objects are there under the obj_class annotation.

Which number represents which object?
is 0 human?

while running task 4b. I am getting following error:

Unrecognized function or variable 'clearMOTMex'. 
Error in CLEAR_MOT_HUN (line 50)
[mme, c, fp, m, g, d, alltracked, allfalsepos] = clearMOTMex(gtMat, resMat, threshold, world, VERBOSE);

Error in classEval (line 11)
        [metsCLEAR, ~, additionalInfo] = CLEAR_MOT_HUN(gtdata, resdata, threshold, world);

Error in evaluateTrackB (line 88)
    tendMets = classEval(gtsortdata, ressortdata, allMets, ind, evalClassSet, sequenceName);

Error in evalMOT (line 32)
     [tendallMets, allresult] = evaluateTrackB(seqPath, resPath, gtPath, evalClassSet);
 
>> 

For example, if the original mAP for Task4a is 35, after I duplicate the content of each video's txt file for 10 times, the mAP increases up to 65.

original txt:        after duplicated:
1,                            1,
2,                            2,
3,                            3,
                              1,
                              2,
                              3,
                              ...

The bug lies in Line 182 - Line 204 in evaluateTrackA.m.

...
num_matched = 0;
num_total = length(union(img_ids, gt_img_ids));
for j = 1:num_obj
        ...
        if(ov >= gt_thr(k))
            num_matched = num_matched + 1;
        end
    end
end
ov = num_matched / num_total;
...

If the result txt is duplicated 10 times, num_matched will also be increased by 10 times, while num_total won't get larger since num_total = length(union(img_ids, gt_img_ids)) is a set operation. Thus, ov becomes larger, which leads to the improvement of mAP.

In order to ensure the fairness of the challenge, an assertion should be added here to fix the bug.

I am trying to evaluate some trackers on the VisDrone dataset.

• https://github.com/mikel-brostrom/Yolov5_DeepSort_Pytorch
• https://github.com/ifzhang/FairMOT

I am using their result text files to run evalMOT.m on, with the runTrackerAllClass line commented off.
Example of outcome using YoloV5_DeepSORT on VisDrone2019-MOT-test-dev:

1. Does the result seems correct?
2. What is the difference between the 2 results ( I am guessing all 10 classes result on top and pedestrians only result below? )

when I run evalMOT.m to eval on test-dev set, I get an error:
tracking the sequence 1/17...
错误使用 load
无法读取文件 'E:\VisDrone-MOT\VisDrone2019-MOT-test-dev\uav0000009_03358_v.txt'。没有此类文件或目录。

出错 runTrackerAllClass (第 12 行)
det = load(fullfile(detPath, [nameSeqs{idSeq} '.txt']));

出错 evalMOT (第 26 行)
runTrackerAllClass(isSeqDisplay, isNMS, detPath, resPath, seqPath, evalClassSet, threSet, nmsThre, trackerName);

Did any one knows this problem? Thank you.

Hi, my code settings are as shown in the figure. Do you know why all the evaluation results are 0 ?