Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
In general, the article is very well written, although at certain places the readers will be easily lost because of some convoluted explanations of technical details, e.g.; - why edges are divided in two types: internal and external? - Are them used differently in benchmarking? - Are those non-recurring traffic-related incidents collected used in benchmarking? - What can they be used for and how? - etc. On the other hand, the paper is well introduced and motivated, raising very challenging problems which may be studied using STREETS. Also, it is very well connected with the literature: section two is a fine and complete summary of existing literature on traffic and vehicle datasets. However, experiments (benchmark evaluation) are not so convincing. The experimental section should be improved, as there is a clear lack of insight as to how “STREET” may be useful to validate and develop new ML models in the area, or how it could be used to guide future work using the dataset. Authors only show two simple baselines addressing one task, which is clearly insufficient to draw any conclusions. Much more tasks and SOTA baselines should have been included to demonstrate the potential of the dataset. Authors argue that this is due to the lack of space, however they could have reduced the description of the procedures and algorithms for data collection, or they could have used the appendix. ------------------- I am changing from "an OK submission, but not good enough" to "marginally above acceptance threshold" after reading the response an the rest of reviews. I appreciate the additional results provided by the authors.
This paper describes an original contribution of a large dataset of street traffic images. The quality of the dataset is excellent and the manuscript is written very clearly, with a great description of the background and prior work. Overall I find this dataset to be a very significant contribution to science.
### Edit: Dear authors, Thank you for your rebuttal. I encourage you to continue to expand the baselines, include the details about the annotations in the appendix of the paper, and work on providing a well-documented code release. I trust you will do all of these things. Though I'm impressed with the rebuttal, I will probably not change my score since I think it is already quite high. I do however think the paper should be accepted. Thanks for your contribution! Best, R3 ### Originality -STREETS differs from previous work in several ways: --The dataset is collected from a camera network with a graph-based structure, and the relationship between cameras is available --The dataset focuses on the suburban setting --The dataset accumulates temporal traffic data from multiple intersections Overall, the authors have clearly explained how STREETS is different from prior work and have explicitly developed STREETS to address shortcomings in this work. Quality -The process for collecting the dataset is straightforward, but clearly described, which contributes to the technical merit of the paper. -The authors contribute a rich amount of metadata, including: --Timestamp, location, view for each image --Hand-labeled masks for cars for the inbound and outbound sides of the roadway at each camera view --Number of traffic lanes at each camera, location of camera, and distance between cameras for recontstructing the traffic graph --Incident data from local transportation authorities -The authors evaluate state of the art object detection networks (Mask-RCNN) for counting vehicles in crowded scenes. -The authors include baseline evaluations on the STREETS dataset for the single-step traffic prediction problem Clarity -The submission is clearly written and well organized. I found that the authors did a particularly good job of motivating the problem and I enjoyed the discussion of the tradeoffs of various traffic forecasting data sources in the introduction. The related work was also particularly well written and informative. I am not an expert in traffic forecasting, but after reading the paper, I understood the problem the researchers were trying to address and had enough context to know why their contribution is significant. Significance -The STREETS dataset addresses many shortcomings found in prior work. Many more researchers in the NeurIPS community should be building datasets like this and I applaud the authors for their efforts. Additionally, many researchers will benefit from this dataset and the work will inspire the development of new methods that will help advance current technology for a meaningful problem.