Benchmark based on MOT records relevant evaluation indicators and their calculation methods and significance

MOTA ( Multiple Object Tracking Accuracy)

MOTA calculates the matching situation of all frames, t is the number of frames, FN is False Negative, FP is False Positive, IDSW is ID Switch, GT is the number of Ground Truth objects, calculates the number of missed and False checks of all frames and the change of ID. ∑tFNt∑tGTt\frac{\sum_t{FN_t}}{\sum_t{GT_t}}∑tGTt∑tFNt, ∑ tFPt ∑ tGTt \ frac {\ sum_t {FP_t}} {\ sum_t {GT_t}} ∑ tGTt ∑ tFPt. ∑tIDSWt∑tGTt\frac{\sum_t{IDSW_t}}{\sum_t{GT_t}}∑tGTt∑tIDSWt where FN and FP are used to observe the performance of detector, while IDSW is used to observe the performance of detector and tracker simultaneously.

MOTP (Multiple Object Tracking Precision)

Where dis distance, refers to the difference between the coordinate box detected between target I and the corresponding GT in frame T. Iou or Euclidean distance is generally adopted. Ct is the number of targets matched in frame t.

IDF1(Identification F-Score)

In tracing, the calculation rules of IDTP, IDFP and IDFN are as follows

In the figure, the first frame of each id change is FP, if it is not detected, it is FN, and if the id of the previous frame is the same, it is FP.

MT (Mostly Tracked)

Satisfy the Ground Truth to match successful tracks at least 80% of the time, accounting for the proportion of all tracking targets. The matching success here does not consider whether THE ID changes, but mainly measures the performance of the detector

ML (Mostly Lost)

The proportion of tracks that meet Ground Truth only match successfully less than 20% of the time in all tracking targets.

Rcll (Recall)

Measure detector performance by dividing the number of correct targets detected in the detector by the total number.


R c l l = ∑ t T P t ∑ t G T t Rcll=\frac{\sum_tTP_t}{\sum_tGT_t}

HOTA(Higher Order Tracking Accuracy)

! [insert picture description here] (HTTP: / / https://p3-juejin.byteimg.com/tos-cn-i-k3u1fbpfcp/9332d44b97f94c91b964ee02a17e5e0b~tplv-k3u1fbpfcp-zoom-1.im age)Copy the code

TPA(True Positive Associations),FPA(False Negative Associations), FNA(False Negative Associations)

According to the above, we can see that HOTA quantifies the adhesion degree of the sequence.

TPA_c: for C in the figure, it is the ID of all the tracks of C that match the actual tracks of C

FPA_c: There is no target matching the actual trajectory of C among all the trajectories of ID when c is predicted in the figure

FNA_c: There is no matching target between the real and predicted trajectories of C in the figure