About classification results about occluded_vehicle_acoustic_detection HOT 2 CLOSED

Hi, thanks for your question!

If you ask how we trained the model, then the answer is that for our dataset (Table I in the paper) we only extract training/validation/test samples when the vehicle is either still occluded behind the first (e.g. left corner) while approaching, or within line-of-sight. No samples are extracted when the vehicle disappeared again behind the other (e.g. right) corner. That means that we do not explicitly train the model to give a particular response after the vehicle has crossed the view, and this situation is also not explicitly evaluated in the initial ablation study (Table II).

However, we do see some occurrence of this effect in our online experiments where the trained model is applied in a sliding window. E.g. see Figure 8(a), where you can see that after the crossing, when the probability of "front" goes down, indeed the probability of "right" increases as one might expect. Vice versa the probability of "left" increases in Figure 8(b) after the vehicle moved right to left. So, while we did not explicitly train for this, in practice we can observe that our model will indeed recognize that the vehicle has moved behind the other corner when applied in an online filter. It appears that the reflection patterns that we classify appear sufficiently similar, even though the vehicle is moving away now.

Hope this answers your question.

from occluded_vehicle_acoustic_detection.

Thanks for your prompt answer.

First of all, there have been some methods using millimeter wave radar and computer vision methods to detect non-line-of-sight vehicles. Your idea of using multi-channel acoustic information for non-line-of-sight vehicle detection is amazing.

And your explanation is very detailed, you also gives the possible reason corresponding to the classification results. I agree with you very much. Due to the similarity of reflection patterns, the model classifies the vehicles that are moving away again. The question I asked was well answered by your answer.

Thank you for sharing your code on GitHub.

from occluded_vehicle_acoustic_detection.