Unveiling Temporal Telltales: Are Unconditional Video Generation Models Implicitly Encoding Temporal Information?

Thank you again for having taken the time to thoroughly evaluate our submitted manuscript.

In reaction to this, we went at great lengths to thoroughly address and accommodate every single one of your comments.

We would very much appreciate you going through our responses to your comments, as well as the revised version. We hope that you are able to reconsider your evaluation and potentially provide additional feedback for a final version of the paper.

Many thanks in advance for your time and efforts.

---

Thank you for pointing out the part where readers could get confused. We have de-emphasized the human angle and focused more on the fact that temporal location is much easier to classify in the generated video than real in the revised version.

In the result section we introduced a metric that should be considered when evaluating the generated videos since real-world videos are not temporally classifiable while generation tasks aim at generating realistic videos, not videos that can be temporally detected easily.

Lastly, there is no clear answer for what features human spatiotemporal perceptually view. Generally speaking, when humans are presented with a video, we focus more on making a video that could contextually make sense. In other words, we look for the relationships between the frames. On the other hand, through our findings, temporal classifiers are not acting in the same way.

It is great to hear that some of the initial concerns have been dismissed after our first response.

Regarding the choice of features for generating naturalistic videos, we emphasize the significance of considering 'viewing time as a continuous signal.' In the context of this paper, the removal of temporal information proves crucial for generating realistic videos along the temporal axis. We assert that current unconditional video generation models primarily prioritize visual appeal without delving into the fundamental challenges of video generation.

In this paper, we contend that the lack of realism in current unconditional video generation stems from the encoding of temporal information in each frame. As the author, I emphasize that the spatiotemporal feature crucial for generating naturalistic videos is the temporal information derived from real videos. This information can be extracted through a simple network or an RNN, but it must align with real-world statistics and characteristics. As mentioned in the original manuscript, we believe that treating temporal information/features as a continuous signal, as demonstrated in previous research [1], is instrumental in addressing implicit temporal encoding.

[1] Skorokhodov, Ivan, Sergey Tulyakov, and Mohamed Elhoseiny. "Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.

Thank you again for having taken the time to thoroughly evaluate our submitted manuscript.

In reaction to this, we went at great lengths to thoroughly address and accommodate every single one of your comments.

We would very much appreciate you going through our responses to your comments, as well as the revised version. We hope that you are able to reconsider your evaluation and potentially provide additional feedback for a final version of the paper.

Many thanks in advance for your time and efforts.

---

(Q1 & W1-1) Why not use MoCoGAN-HD and StyleGAN-V in the preliminary experiment? This would avoid reproducing MoCoGAN-256. / Experiments should encompass multiple baselines to bolster the claim that current unconditional video generation methods implicitly encode temporal information.

(Response – Q1 & W1-1) The results in the preliminary section can be extended if we bring MoCoGAN-HD and StyleGAN-V results in the experiment section (Table. 2, 3, 4). However, we only presented the MoCoGAN results to avoid redundancy. Nevertheless, we agree that it would be better bolstered if we tested on diverse models (i.e., Diffusion-based model). As suggested by another reviewer, we have tested on recent diffusion-based unconditional video generation model to show the problem of implicit temporal encoding does not occur only in the GAN-based models but also in the diffusion-based model. Similar to the GAN-based model, the diffusion-based model also showed the existence of the problem by being able to temporally classify the generated videos. We have added the section in the revised appendix for this experiment (45.55% in Sky-Timelapse and 19.66% in the UCF-101).

(W1-2&3 & Q2) The constructed dataset used for training the temporal classifier appears overly homogenous since the clips are sampled from a single video clip of FaceForensics, Sky–Timelapse, or UCF-101. / The labels for certain frames seem less meaningful due to the frame's presence in various clips at different positions, arising from the repeated random sampling during the construction of the temporal classification dataset. / Can you visualize some real and generated video frames used in the preliminary experiment? This could enhance clarity.

(Response – W1-2&3 & Q2) As we fix the content noise vector while varying the motion noise vector, the generated videos consist of frames that look similar to another frame located in another temporal location and are also overly homogenous. With these generated videos, the CNNs are able to classify the temporal location of the frames better than the real-world videos. To mitigate the same dataset construction, we have meticulously designed the real-world dataset for temporal classification. We have added some samples for temporal classification in the revised appendix.

(W2-1 & 2) Figure 2 lacks an introduction of f_{temp} in the caption, and the caption references an ImageNet pre-trained model not presented in the figure. / An invalid figure reference in the last paragraph of Section 3 suggests a missing figure in the manuscript.

(Response – W2-1&2) Thank you for identifying places where additional explanations were needed and for pointing out typos. We have added the explanation in Figure 2 for the f_{temp} in the revised version. The ImageNet pre-trained model that we utilized is inside the torchvision package provided by PyTorch and there is no exact reference for this. However, we did add that we utilized ImageNet pre-trained ResNet-18 in the caption of Figure 2. The invalid figure reference in the last paragraph of Section 3 was supposed to refer to Figure 2 in the manuscript, we have corrected the reference in the revised version.

(W3) The training quality of the reproduced MoCoGAN-256 appears suboptimal. Tables 2, 3, and 4 reveal an extremely high FVD value for MoCoGAN with 256x256 compared to the other high-resolution video generation baselines.

(Response – W3) MoCoGAN was the first unconditional video generation model that disentangled the content and motion noise vector for generating videos. It is important to note that the primary focus of MoCoGAN was not centered on generating high-resolution videos such as 256x256 but on generating 64x64 videos. In the MoCoGAN paper, they experimented on the Weizmann Action dataset and UCF-101 resized to 64x64. Therefore, the MoCoGAN-256 could seem suboptimal when compared to MoCoGAN-HD and StyleGAN-V, the methods aimed at generating 256x256 videos.

(Q5) Can you provide some failure cases of video generation with GRL?

(Response – Q5) We acknowledge the importance of presenting failure cases for advancing research in this field. However, we have not observed any failure cases while training. However, a potential failure scenario arises when the temporal classifier struggles to classify temporal location during training. Specifically, if the generated videos lack temporal information (i.e. if the generator is not implicitly encoding the temporal information) it could result in 1) a misclassification of videos in terms of temporal location and consequently 2) a failure of the Gradient Reversal Layer (GRL) to fulfill its intended function.

(W4) The validity of the negative gradient provided by the temporal classifier during training needs reconsideration. The temporal classifier is trained using frames from different videos, which is different from the practice in the preliminary experiment and the evaluation stage that constrains the training frames to have the same content.

(Response – W4) A batch of generated videos for training the temporal classifier, different from the videos utilized for training the discriminator, is generated with fixed content noise as noted by the reviewer. We added an additional explanation in the revised version.

(W5) The absence of recent diffusion models for video generation (Luo et al., 2023; Yu et al., 2023; Harvey et al., 2022) in the experiments diminishes the contribution of this work.

(Response – W5) We admit that the problem has to be general across different backbones (GAN-based model and Diffusion-based model). From the suggested methods, we have tested PVDM[1] as they have an open-source implementation with the author-provided checkpoints. Due to the time constraint of the discussion period, we could not train the PVDM and PVDM + ours to demonstrate the effectiveness of our method. However, we did an experiment to see if the implicit temporal encoding is also occurring in the diffusion-based model. For the experiment, we generated the videos in Sky-Timelapse and UCF-101 with the checkpoint provided by the author. Consistent with our experiments in the manuscript, we have fixed the content noise vector while varying the motion noise vector, and maintain other experimental settings identical to the original experiment. The table below shows the temporal classification accuracy of the generated videos. The results demonstrate that implicit temporal encoding also occurs in the diffusion-based model. We have added the section in the appendix in the revised version with the visual samples of the generated videos that were used in the experiment.

(W6) Additional insights on designing architectures that "do not necessitate classification" in the discussion section would be beneficial.

(Response – W6) We sincerely appreciate the reviewer for raising this point. This is definitely a go-to way for future research. As stated in the discussion section, employing a classification model may pose challenges when generating extensively long videos (i.e., 1-hour or longer videos). We may try to make the classification model to learn the relative temporal location, however, this may be suboptimal or even not work properly. Unfortunately, we do not currently have any insights on architecture that do not necessitate classification. If possible, it would be a great honor to hear back the thoughts on the potential architecture construction.

(Q3) What is the reason for using distinct videos for evaluating temporal accuracy and FVD computation? Why not use the temporal classifier from the training process to evaluate temporal classification accuracy?

(Response – Q3) The FVD metric measures the feature distance similarity between the real-world videos and generated videos. Recognizing the suboptimal quality of early-stage generated videos, we employed a fully trained unconditional video generation model for generating images for measuring FVD and temporal accuracy. As for the temporal classifier, we did not use the one from the training as the reviewer stated. While the temporal classifier was designed to accurately classify the temporal location of frames during training, the inclusion of the Gradient Reversal Layer (GRL) before the temporal classifier may have harmed its classification ability. To ensure precise measurement of the temporal accuracy we re-trained the temporal classification model. To make a fully functional temporal classification, one would need a sufficient amount of data. In this work, we generated 2,048\times0.8 videos for training.

(Q4) Can you provide qualitative comparisons on UCF101? This would provide additional insights.

(Response – Q4) The quality of the generated UCF-101 videos hampers from comparing the results visually. Prior works also show the Sky-Timelapse and FaceForensics for qualitative comparison as done in our paper. However, we added some UCF-101 results in the revised appendix to provide additional insights to the reviewer.

[1] Yu, Sihyun, et al. "Video probabilistic diffusion models in projected latent space." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.

Thank you for your comprehensive feedback. Each point raised in the initial comments has been carefully addressed, with a focus on understanding the underlying concerns. However, it appears that the concerns persist, particularly with regard to the temporal classifier. Further clarification on the specific aspects perceived as unaddressed would be greatly appreciated, enabling a more targeted and effective response.

Furthermore, I acknowledge your observation about the improvement in the FVD score without a corresponding enhancement in the visual quality of the generated video. Alongside the explanation in the original manuscript, I would like to underscore that the paper addresses a specific issue in the video generation task, namely the encoding of temporal information. Consequently, a simple neural network can temporally classify the generated videos, a phenomenon we acknowledge as unconventional. To mitigate this temporal encoding, our method introduces only the Gradient Reversal Layer (GRL) with the temporal classifier during video generator training and nothing more. The outcome is an improved FVD score while maintaining visual quality equal to that of the original works. It is essential to assert that our method does not contain any components designed to enhance video generation in terms of video quality and/or FVD score. Instead, our argument centers around the alignment of generated videos with real-world characteristics, justified by the observed FVD score improvement.

Adding to the reviewer's concern about the visual results, it's pertinent to note that developments in image and video generation exhibit similar trends. In image generation, the focus has shifted from creating realistic images to exploring applications like style transfer, as the difference in visual quality became less meaningful. (The visual quality was mostly indistinguishable.) Similarly, in video generation tasks, the evolution moved from generating realistic videos to realistic temporal information [2, 3, 4]. As these tasks began leveraging well-trained image generation models [1, 3], the visual differences became less significant. In line with this, our emphasis on addressing the problem in current temporal encoding in video generation is reflected only through the FVD score not by the visual enhancement, however, we believe that this research will pave the way for future video generation tasks as it, for the first time, revealed the problem in video generation task.

[1] Tian, Yu, et al. "A Good Image Generator Is What You Need for High-Resolution Video Synthesis." International Conference on Learning Representations. 2020.

[2] Yu, Sihyun, et al. "Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks." International Conference on Learning Representations. 2021.

[3] Skorokhodov, Ivan, Sergey Tulyakov, and Mohamed Elhoseiny. "Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.

[4] Wang, Yuhan, Liming Jiang, and Chen Change Loy. "StyleInV: A Temporal Style Modulated Inversion Network for Unconditional Video Generation." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.

Thank you again for having taken the time to thoroughly evaluate our submitted manuscript.

In reaction to this, we went at great lengths to thoroughly address and accommodate every single one of your comments.

We would very much appreciate you going through our responses to your comments, as well as the revised version. We hope that you are able to reconsider your evaluation and potentially provide additional feedback for a final version of the paper.

Many thanks in advance for your time and efforts.

---

(W1) Real-world video image classification setting

(Response – W1) We apologize for any confusion resulting from the lack of clarity in presenting the real-world video image classification setting. As mentioned by the reviewer, we recognized the importance of temporal classification settings between the real-world videos and the generated videos to be as same as possible. To address this, we constructed the real-world video dataset for temporal classification with the following steps. 1) We randomly choose a single content category, 2) We fix the 0th frame as the first frame for all 2,048 videos, and finally 3) We randomly sample the rest 15 frames from the same video. We have revised the paper by adding an explanation of the real-world temporal classification setting in the preliminary section. We believe that the reviewer’s concern that the ‘first frame of one short video may be the last frame of another video’ is resolved through step #2. As for the training of the real-world videos with longer epochs, we saw that training accuracy converged around 120 epochs and selected 150 epochs for temporal classification training. Therefore, we firmly state that the longer training on the real-world videos does not increase the evaluation accuracy or the training accuracy.

(W2) Multiple usage of GRL and its meaning

(Response – W2) Thank you for mentioning the part that may confuse the readers but is essential for understanding our paper. GRL implies the Gradient Reversal Layer in all of its uses in our paper. The main function of GRL is to reverse the gradient during the backpropagation process but does not change the features in the forward process. During training, the GRL backpropagates -1{\times}gradient to the generation model while backpropagating the original gradient to the temporal classifier. This makes the generation model to explicitly not capture the temporal information while the temporal classifier tries to identify the temporal features and classify them to the correct temporal location.

(W3) Visual results

(Response – W3) We indeed agree with the reviewer in viewing the visual quality over FVD in the generation task. However, as mentioned by the reviewer, the core message of the paper is the problem of implicit temporal encoding in generated videos. Additionally, we emphasize that our proposed method does not have any module or loss term to directly benefit the visual quality of the generated videos, rather, we have neglected the implicit temporal encoding when generating videos. To further demonstrate that the proposed method does not harm the visual quality of the original works while achieving higher FVD just by neglecting the temporal information encoding, we have added a few samples of generated videos in the revised appendix.

(Additional Comments)

(Response – Additional Comments) Thank you for catching the typos, we have revised the paper accordingly.

Thank you again for having taken the time to thoroughly evaluate our submitted manuscript.

In reaction to this, we went at great lengths to thoroughly address and accommodate every single one of your comments.

We would very much appreciate you going through our responses to your comments, as well as the revised version. We hope that you are able to reconsider your evaluation and potentially provide additional feedback for a final version of the paper.

Many thanks in advance for your time and efforts.

---

(W1) Lack of comparison with recent methods, e.g., [1, 2, 3]. These methods seem to have much lower FVD than the proposed methods on UCF101 dataset.

(Response – W1) For comparison, we have selected a recent GAN-based model in the unconditional video generation field that incorporates separate content and motion noise vectors. Also, we highlight that we are not claiming to have achieved state-of-the-art performance compared to recent works. Rather, we emphasize that by neglecting implicitly encoded temporal information, we obtain a better FVD score while maintaining similar qualitative video generation results. However, we admit that experimenting with diffusion-based unconditional video generation models enriches our paper by indicating that implicit temporal encoding occurs not only with the GAN-based models but also with diffusion-based models. Due to the time constraint of the discussion period, we were unable to train the diffusion-based model and diffusion-based model + ours. Consequently, we focused on testing whether the generated videos from the trained diffusion-based model are temporally classifiable or not. We searched for a diffusion-based model that has open-source implementation and the checkpoint provided by the author. We were not able to find the right fit in the reviewer’s suggested methods therefore we present the experiment done on another reviewer’s suggested method, PVDM[1]. In line with our original experiments, we have fixed the content noise and varied the motion noise in PVDM. The generated videos of Sky-Timelapse and UCF-101 with PVDM do have higher temporal classification accuracy when compared to the real videos in Sky-Timelapse and UCF-101 as shown in the below table.

Although the difference in test accuracy of UCF-101 may seem negligible, we argue that an improvement in the quality of the generated UCF-101 videos could correspond to an increase in temporal accuracy. We added qualitative and quantitative results in the appendix of the revised version as this could give insights for research in diffusion-based generation models.

(W2) Marginal performance improvement. The proposed method achieves slightly better performance than its baseline, i.e., MoCoGAN, on various datasets. However, the margin is quite small, e.g., FVD of 2539 (MoCoGAN) v.s. FVD of 2360 (proposed method). Such FVD improvement may not be sufficient to convince readers that visual quality of videos generated by the proposed method is better than that of MoCoGAN.

(Response – W2) The FVD score may not appear as appealing when considering a method designed to generate higher-quality videos, incorporating novel loss terms or modules with the goal of enhancing the video quality. However, our emphasis in this paper is on introducing a novel problem in current unconditional video generation models. Experimentally we demonstrated that neglecting this problem contributes to an improvement in the performance of video generation in terms of FVD score.

(W3 & Q1) I am not able to see why encoding temporal information in generated videos is a major problem that prevents GAN-based methods to generate high quality videos. / Why encoding temporal information in generated videos is a major problem that prevents GAN-based methods to generate high quality videos?

(Response – W3 & Q1) As implied in the manuscript, our main statement is that videos generated with current unconditional video generation methods have temporal information encoded as the reviewer said and our method ensures temporal information to be not encoded in generated videos. More importantly, we are not proposing additional loss terms or modules aiming at generating better quality videos. We demonstrated that by neglecting such temporal information encoding when training the generation model, we achieved better FVD scores while showing similar visual quality with the original works as shown in Figure 3 (FaceForensics and Sky-Timelapse) and in the revised appendix (UCF-101).

[1] Yu, Sihyun, et al. "Video probabilistic diffusion models in projected latent space." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.

(W4) Minor issues (1) There are "??" in this paper. (2) The authors claim that they investigate "meaning of ‘realness’ in the video generation models" in the abstract. However, I am having a hard time finding how the meaning of 'realness' connects to the proposed method.

(Response – W4) Thank you for catching the minor mistake, the correct reference can be seen in the revised version. While acknowledging the challenge of precisely defining realness, we attempted to characterize it as video possessing similar attributes to those in real-world videos and with how humans perceive the temporal location within videos. To quantify this aspect, the commonly used FVD metric could be a suitable choice. However, we propose considering a temporal classification metric as well. This metric aligns with human behavior, connecting the way humans perceive videos and classify temporal location with the notion of realness.

We appreciate the thorough feedback and thoughtful consideration of our latest results. Our primary goal of this paper is to demonstrate that removing the implicit encoding of temporal information leads to a decrease in FVD scores. While the evidence supports this argument, we understand your concern about the modest improvement in FVD not being convincing enough. We recognize that the standard expectation is for FVD score and visual quality to improve concurrently in video generation tasks. In our case, FVD dropped while visual quality remained consistent. We interpret this as the removal of temporal information enhancing the realism of the generated videos in the feature space without impacting the image space.

Adding to the reviewer's concern about the visual results, it's pertinent to note that developments in image and video generation exhibit similar trends. In image generation, the focus has shifted from creating realistic images to exploring applications like style transfer, as the difference in visual quality became less meaningful. (The visual quality was mostly indistinguishable.) Similarly, in video generation tasks, the evolution moved from generating realistic videos to realistic temporal information [2, 3, 4]. As these tasks began leveraging well-trained image generation models [1, 3], the visual differences became less significant. In line with this, our emphasis on addressing the problem in current temporal encoding in video generation is reflected only through the FVD score not by the visual enhancement, however, we believe that this research will pave the way for future video generation tasks as it, for the first time, revealed the problem in video generation task.

[1] Tian, Yu, et al. "A Good Image Generator Is What You Need for High-Resolution Video Synthesis." International Conference on Learning Representations. 2020.

[2] Yu, Sihyun, et al. "Generating Videos with Dynamics-aware Implicit Generative Adversarial Networks." International Conference on Learning Representations. 2021.

[3] Skorokhodov, Ivan, Sergey Tulyakov, and Mohamed Elhoseiny. "Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.

[4] Wang, Yuhan, Liming Jiang, and Chen Change Loy. "StyleInV: A Temporal Style Modulated Inversion Network for Unconditional Video Generation." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.

Thank you for the detailed reply and the new results.

While this paper identifies an interesting potential problem of existing GAN-based video generation models for the first time, some of questions are not well addressed, e.g., W2 and W3. It seems to me that the goal is to show that the temporal information harms FVD but not visual quality. While there are evidence that supports this argument, the small FVD improvements is not convincing enough. More importantly, the goal of video generation is not to generate videos that have low FVD but to generate videos that have great visual quality. I am not sure how the findings of this paper help researchers achieve this goal.

Thank you again for having taken the time to thoroughly evaluate our submitted manuscript.

In reaction to this, we went at great lengths to thoroughly address and accommodate every single one of your comments.

We would very much appreciate you going through our responses to your comments, as well as the revised version. We hope that you are able to reconsider your evaluation and potentially provide additional feedback for a final version of the paper.

Many thanks in advance for your time and efforts.

---

(W1) It is not clear why encoding the temporal information inside the frame would lead to the video quality degradation. Some empirical / theoretical explanation could be useful to provide further insights.

(Response - W1) The temporal information inside the frame does not lead to visual quality degradation of the video. To put it in another way, the temporal information does affect the FVD but not the visible quality of the video as can be seen in Figure 3 (FaceForensics and Sky-Timelapse) and in the revised appendix (UCF-101). In terms of FVD, we hypothesize that it could degrade due to temporal information encoding as each frame from a generated video has additional information that is not essential/needed in case of the real-world videos.

(W2) The argument that since CNN is inspired from humans, then they should not be able to detect the temporal signal embedded in the generated frames is not necessary. The main point of the paper tries to show that the generated videos have clear temporal information inside a single frame. The author could raise less confusion if framing the CNN detector is a simple quantitative tool to detect the time information embedded inside the video frame.

(Response - W2) Thank you for addressing the concern that could potentially hinder readers from understanding our paper’s core message. The initial purpose of stating ‘CNN is inspired from humans’ was to illustrate why the problem of implicit temporal information in the generated videos is an incomprehensible phenomenon. However, as we agree that argument could be minimized and revised the paper accordingly.

(W3) The paper would become more appealing if framing it as a method against fake generated video and find applications in face swapping detection. Adding some experiments in this domain could make the impact broader.

(Response – W3) Thank you for pointing out an application of the temporal classification metric in another field. To assess the effectiveness of the proposed temporal classification metric in detecting deepfake videos, we utilized the Celeb-DF[1] and Celeb-DF-v2[1] datasets which are commonly used in the field of deepfake detection. Each dataset consists of real and synthesized videos where the synthesized videos are made by face swapping the real video with another face identity with an auto-encoder. For a fair comparison between real and synthesized videos, we first randomly select a video from real videos. Then, we select the same video with another face identity in the synthesized video. In line with our original experiments, we fix the first frame with the 0th frame from each video and randomly sample the rest 15 frames from the video. We note that the 15 frames are of the same index, meaning that they are from the same temporal location. The temporal classification accuracy of the synthesized video is around 1.07%p higher than real video results. As the difference is not as notable as the results in the main table (with generated and real videos), the temporal classification metric cannot be directly used for detecting deepfake videos. With curiosity we have also experimented with face-cropped versions of each video, considering that synthesized videos are generated solely by face-swapping the original video. We obtained face-cropped versions utilizing the Haar Cascade algorithm. The result followed a similar trend with the initial experiment showing around 0.92%p difference. We hypothesize that the small temporal accuracy difference comes from not generating entire videos but rather only detecting and swapping faces, thereby leaving the temporal characteristics of the real video unchanged. However, as the extended experiment on the deepfake videos may give a pathway to future research in deepfake detection, we added a section in the appendix in the revised version.

(Q1) The authors mentioned about the human study but there is no explicit section to document the details of how to conduct the human study.

(Response – Q1) The user study mentioned in the manuscript is to emphasize the usage of the user study in the prior works. We have not conducted any human studies for our experiment and thus do not have a section in the original paper.

[1] Li, Yuezun, et al. "Celeb-df: A large-scale challenging dataset for deepfake forensics." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2020.