NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:4701
Title:Seeing the Wind: Visual Wind Speed Prediction with a Coupled Convolutional and Recurrent Neural Network

Reviewer 1

[Updated after the author feedback] =Summary= The goal of the paper is to devise a mechanism to predict wind speed based on the 2-second video snippets on a checkered flag. A resnet-18 is used as a feature extractor for each frame. The timeseries of features is then fed to an LSTM to predict the wind speed. A specialized dataset is also collected for the aforementioned task. The paper attempts an interesting problem. The idea of using visual signals to predict wind speed is indeed quite clever. However, in the current state, the paper needs more work before it is ready for publication. On the plus side, the collected dataset is quite useful for the boarder ML community. On the negative side, the paper neither introduces a theoretical or a methodological novelty, nor does it show striking empirical evidence. The empirical results are somewhat underwhelming and certain choices and intuitions are not explained very well. Overall, this is a promising research direction, and the authors are encouraged to pursue it by refining the writing and expanding the experimental setup. Please see the suggestions for improvements below. =Originality= Medium. The paper applies well-known models to a specific application. The gathered dataset however can prove to be quite useful. =Quality= Low. The model choices could be made in a more careful manner. For example, it is not clear why only recurrent networks would need to be used here. Please see detailed comments below. =Clarity= Medium. The content is generally accessible. =Significance= Low-Medium. The data collection and curation procedure used in the paper is quite thorough and could serve as a nice guideline for ML practitioners. However, on the negative side, the paper is very application specific and the results are somewhat underwhelming. Given the limited scope of the empirical validation, the paper is unlikely to have significant impact on the broader ML community. = Specific comments and suggestions for improvement = - The rebuttal was helpful in clarifying the operational mode of the proposed method, that is, once the method training is done, one can use the pre-existing structures e.g., trees and flags (as opposed to planting custom flags at each point of interest) to measure wind speed. This is a pretty clever idea and this reviewer highly recommends explicitly mentioning it in the intro. - In relation to the previous point, the performance on test sets (tunnel test set and adjacent flag test set) is a bit underwhelming. Looks like the method is overfitting to the main flag. Perhaps, it would help to have a separate validation flag and fine tune the model hyperparameters (architecture, early stopping) based on performance on this separate validation flag? Most importantly, it would be great to extend the experimental setup to other (non-checkered) flags/trees to empirically show that the idea has the promise to generalize. In summary, the paper can greatly benefit from a more comprehensive empirical section. - The intro mentions that the paper leverages both "physics and machine learning to predict wind speeds". However, the method is completely data-driven (CNN + LSTM). - It is not clear why the LSTM needs to be used in this setting. It is true that LSTMs are the first model that one thinks of when modeling timeseries data, however, for data with fixed length and relatively small (30 timesteps) trajectories, a CNN might work equally well. Additionally, it seems that given the nature of the problem (motion of an object over time), it seems that the self-attention mechanisms might be well suited for this task ( Regarding the answer in rebuttal to this point, this reviewer is not convinced that once can use the LSTM trained on 15 fps to make predictions on 30 fps. - It section 3, it seems like bins containing disproportionately large data are down-sampled. Rather than discarding data, why not upsample the sparse bins? - While this reviewer is not an expert on wind speed prediction, it sounds like taking 1-minute speed averages might be too noisy, that is, wind speed might change a lot during this time. It might be possible to check this variance from the data. The question that remains after the rebuttal is that how precisely was the number of 1-minute chosen? How would the takeaways change if one considered 30-second averages? - How were the training and validation sets separated? If one considers a 1-minute interval, then there are thirty consecutive 2-second intervals in this time period. However, all of these intervals might potentially have the same wind speed value (since it is a one minute average). Giving this overlap, would it make sense to split the training and validation data such that it does not overlap in these 1-minute intervals? - It would be great to mention if / how the analysis of 5.1.1 gets affected by the fact that the ground truth wind speeds considered in the paper are 1-minute averages.

Reviewer 2

This paper uses a coupled CNN and RNN architecture to predict wind speeds from 2-second clips of flags (specifically, observing their flow-structure interactions). The authors construct a dataset of checkered flags in three environments: (1) at a field site in Lancaster, CA, (2) at the same field site and with the same type of flag, but 3m away from the original flag, (3) in a wind tunnel, with a smaller-sized flag, and with monochrome camera videos. They then employ their model on this dataset, using a CNN to do feature extraction and then using an RNN to predict the wind speed associated with a 2 second (30 frame) flag video. The authors employ mean subtraction on their dataset to remove background features so that their model will better generalize to different conditions. They also contextualize the performance of their model at different wind speed ranges via frequency analysis (of the Nyquist frequency and clip duration-limited frequency) and turbulence analysis (to understand error bars). Originality: To the best of my knowledge, this is the first work that uses ML for wind speed prediction from flags. The closest related work (which the authors cite) is in extracting wind measurements from instrumentation (their instrument of choice here is nontraditional -- a flag) and in video prediction (which has not before been applied to this domain). From an applications perspective, this is therefore certainly a novel contribution (and one that would be useful to practitioners). The model evaluation portion also demonstrates originality, as model predictions are contextualized using frequency and turbulence equations. Quality: The work seems very well-executed, with assumptions and limitations both explained and contexualized in the domain. The model seems to predict wind speeds well on the main and adjacent flag test sets (i.e. the ones at the field site) for a variety of wind speeds, and does okay on the wind tunnel test set (which is not surprising given the different camera type and different flag size). I wish there had been more analysis of why the authors feel predictions for the tunnel test flag set were so uniform. I also have questions about the error bar calculations (see next section of the review). Overall, however, I believe this is a solid, well-executed contribution with good results, and that it poses a lot of challenges for future work to build on. Clarity: The submission is very clear and easy to follow. The description of the related work, dataset, method, model evaluation are all very clear and written in an engaging manner. Given the dataset, I could likely reproduce the authors' approach. Significance: I believe this work is significant from multiple perspectives. First, it provides a novel method for predicting wind speeds using commonly-available "instrumentation." Second, it provides a great case study of domain-specific considerations that are necessary for applying ML in the field -- for instance, previous ML algorithms have overfit to background features, and the authors explicitly account for this by collecting and using a diverse dataset. Third, it provides a dataset that the community can use for further exploration of this problem (though this dataset can and should be expanded upon with a more diverse set of flags and conditions in future work). EDITS AFTER DISCUSSION & AUTHOR FEEDBACK: * Initially, my biggest concern about this paper was that the dataset was not diverse (and I believe this concern was shared by the other reviewers). That said, the author response indicates that the authors have collected additional data (tree canopy movement), and that their model performs well on the tree canopy validation set when trained on tree canopy data. This additional data increases my impression of the value of the dataset presented here, and I do believe the dataset is one of the biggest contributions of this paper. * I am now less convinced that the model generalizes due to the flatness of the test set results (though it's worth noting that most NeurIPS papers evaluate on the equivalent of the validation set -- on which the authors' models perform well -- as opposed to the test set). (I also laud the authors for their honesty in providing corrected results.) Additionally, while I am glad to see that the model performs well on the tree canopy validation set when trained on tree canopies, to understand generalization, it would possibly be more important to see whether e.g. a model trained on the checkered flags generalizes to other flags/the tree, or a model trained on this tree generalizes to other trees. Given my increased impression of the dataset and my slightly decreased impression of generalization, my score remains the same.

Reviewer 3

UPDATED REVIEW: I appreciate the extra experiments and clear explanation of them, and am happy to raise my score. I would have liked to see some discussion of the cost of flags vs. anemometers, but maybe this is in the "specific comments" that the authors say will be incorporated. ================== Review summary: The paper is well-written, the dataset and the experiments done are well explained and careful. I like this paper and definitely want to encourage this line of work, but I am on the fence about whether there is sufficient experimentation here to merit publishing at this stage. Originality: The idea of estimating wind speed from imagery (video) is novel to my knowledge, and in my opinion is the main contribution of the paper. There is nothing novel about the model (CNN feature extractor with an RNN on top), but the authors don't claim there is, and based on the results it seems to perform well. Quality: The paper is well-written, and the dataset and experiments seem to be well described and of good quality. In order to be a really high quality paper though, I would want it to convince me that this idea has practical merit - i.e. at a minimum I would want to see experiments with one other video, e.g. a different type of flag, or a moving tree. Clarity: I found everything very clear; one of the strongest aspects of the paper. Significance: Difficult to judge without more convincing experiments, but it is at least an interesting idea and the dataset is a solid contribution; overall I would say "medium". - what about using auditory information and/or natural language reports of wind speed along /instead of the video? I would think this would be even cheaper than video, and making use of multimodal information would fit with the motivation to use 'existing' data in practical situations (e.g. drone delivery). This is just an idea, not a suggested improvement.