Issues about Mapped Mahalanobis Distance

[fbh16]

Hi Corfyi, many thanks for releasing this impressive work. I have a confusion regarding the construction of the measurement noise covariance matrix R_k in the CMD (Correlated Measurement Distribution) section. Specifically, in relation to the detection noise factor(sigma_m=0.05) which is encountered within the getUVError function in the code, my understanding is that some fine-tuning may be necessary for images/videos with different resolutions. Additionally, when dealing with images or videos that have resolutions differing from the demo, I'm uncertain about how the constraints on pixel errors along the u and v axes in the getUVError function (specifically, the several hyperparameters like u>13, u<2, and v>10) should be adjusted. However, I am not entirely confident in my interpretation, and I would greatly appreciate hearing your perspective on this matter.

[corfyi]

Thank you for your question and interest in our work. Regarding the impact of $\sigma_m$ on tracking performance, we have discussed this in our ablation studies. As shown in Figure 4(b), the influence of $\sigma_m$ is not particularly significant. We consider $\sigma_m$ to be a relative measure that reflects the uncertainty in the detection of target height, which is largely independent of the image resolution.

In our practical implementations, we used a value of $\sigma_m = 0.05$ and found that fine-tuning this parameter is not necessary. Its impact on the tracking performance is negligible compared to the influence of parameters like $w_x$ and $w_y$. Therefore, in scenarios with varying image resolutions, adjusting $\sigma_m$ might not yield substantial changes in the tracking results.

I hope this clarifies your doubts. If you have any further questions or need more information, please feel free to reach out.

[fbh16]

Hi Corfyi,

I have another question I'd like to consult you about. In the paper, the Mahalanobis distance $D$ between the probability distributions of the observations and predictions is calculated to measure their similarity, thereby maintaining the target's ID at different time steps.

My question is, does a smaller $D$ indicate higher similarity between the distributions of the observations and predictions (i.e., representing the same target), or does a larger one indicate higher similarity?

The $D$ is constructed using the covariance and the mean of the distributions of observations and predictions (Eq. 6-Eq. 8 in the paper). Therefore, the larger the residual $\epsilon$ and the residual covariance $S$ between the two distributions is, the larger the $D$ become. So, I lean towards the idea that a smaller $D$ indicates higher similarity.

However, in practical parameter tuning, I found that setting a larger Mahalanobis distance threshold (cost_limit), with reference to the elements in the cost matrix, actually leads to better tracking performance. Conversely, setting a smaller threshold results in more IDsw. https://github.com/corfyi/UCMCTrack/blob/7851a49078a6666725d1bda670f0046fcc0d715f/tracker/ucmc.py#L15

Could you please share your understanding? If there are any errors in my statement, please feel free to point them out. Thanks!