huangqzj
commented
7 months ago @CheeryLeeyy I'm very sorry for negligence in not promptly noticing your issues. We are delighted that you are interested in our work. Please find the point-to-point response to your issues as follows.
1) How does ‘information availability latency’ contribute? The term "Information availability latency" can be simply understood as the time delay between the capture and utilization of information. It is somewhat similar to the concept of AoI (age of information), but not identical. For instance, let's assume at t=0, agent j's sensor captures a frame of traffic data, and at t=total_dealy, agent i utilizes this data to perceive the surrounding obstacles. Here, the value of total_delay represents the "information availability latency’. Intuitively, a shorter total_delay is preferable, as a larger value may adversely affect perception outcomes (obstacle information with a significant delay can be outdated). In our work, we try to employ a neural network to find this threshold (corresponding to enhance_weight).
2) How are the actual values of its components obtained, and does it impact the accuracy of the final object detection? How the transmission rates of various comparison methods, such as Where2comm and V2VNet, are calculated? According to V2X-Vit and other works, we preliminarily determine that "information availability latency’(hereinafter referred to as the total delay) consists of the asynchronous time of the sensor, the computation time of feature extraction, the network transmission time, and the system waiting time. The specific implementations of these values can be seen in the "time_delay_calculation()" function in basedataset.py for all methods (our Select2Col, Where2comm, V2VNet, and V2X-Vit), and the explanation as below. a)Due to the Lidar's perception cycle of 10Hz, both the asynchronous time and the waiting time are random values between 0-100ms. b)The computation time is based on the performance of the computing device we used and is a random value between 20-40ms. c)The transmission time is calculated based on the transmission size and transmission rate. To ensure fair comparison with Where2comm, V2VNet, and V2X-Vit, we used the same transmission size and transmission rate, meaning that we have the same transmission time. The transmission rate is derived from the basic Shannon's theorem. As described in 1), roughly speaking, the shorter the total delay is, the greater the contribution to perception accuracy. Conversely, if the value is too large, it may even have a negative impact.
3) Would imposing constraints on the transmission bandwidth and rates negatively affect their original performance? Yes, it does have an impact. More precisely, the total delay, referred to as "information availability latency," significantly influences perception performance, as observed in our experiments and V2X-Vit. Within our article (see https://arxiv.org/abs/2307.16517), Figures 9-11 reflect the impact of total delay on perception performance. If we decrease the bandwidth or transmission rate while keeping others unchanged, the transmission time will extend, leading to an increase in the total delay, and subsequent deterioration in perception performance.
4) I'm also wondering about the handling of channel noise. What is the principle for selecting this numerical value? Is the channel noise directly added to the feature layer during transmission? And could you provide insights into its implementation? Thank you for your attention to this matter. Existing works, including V2X-Vit, Where2comm, and V2XNet, focus on collaborative perception from a computer vision perspective. To ensure a fair comparison with these methods, we, like them, disregard the impact of noise on the raw transmitted data and only utilize the signal-to-noise ratio (SNR) to calculate the transmission rate. By the way, we are currently involved in another work that emphasizes collaborative perception from the communication perspective. In this new work, we will comprehensively consider the influence of transmission noise and wireless channel for collaborative perception.
Thank you for your interest in our work. If you have any other questions, please do not hesitate to contact us.
CheeryLeeyy
Hi! I came across your remarkable work and have developed a keen interest in it recently. I have some questions regarding the channel modeling parameters and their corresponding code implementation.
Specifically, I am curious about how semantic feature selection is performed based on the bandwidth and channel capacity of the specific communication scenario. How does "information availability latency" contribute, how are the actual values of its components obtained, and does it impact the accuracy of the final object detection? Besides, I'm curious about how the transmission rates of various comparison methods, such as Where2comm and V2VNet, are calculated. Would imposing constraints on the transmission bandwidth and rates negatively affect their original performance?
I'm also wondering about the handling of channel noise. What is the principle for selecting this numerical value? Is the channel noise directly added to the feature layer during transmission? And could you provide insights into its implementation?
As I don't have enough background in communication, these questions may be quite fundamental. I would greatly appreciate any guidance you can provide.