Claims: Although pose is an input to our model, no GT pose labels were used for training.

We will answer the major points below and address all remaining ones in the final version. We leave further discussions on the convergence rate to future works. Eqn.11 in [B] is generic (a superset of most loss engineerings like [3, 29, A]), it uses bi-level We will add a discussion on [3, A, B] in the final version. Q2: Meta sampler has a similar idea to [12,24,27]. CIFAR10-L T); theirs are instance-based and ours is class-based (fewer parameters and simpler optimization landscape).

We thank the reviewers for their valuable comments and recognition of the novelty and results of our method, e . We respond to the major comments below but will address all feedback in our revised version. Proxies are globally learnable "cluster centers" while Clustering [13] directly regards There are actually two types of constraints among proxies in our method, i . "soft" constraint, by encouraging proxies to be close to their anchor samples ( In practice, similar proxies tend to be sufficiently close to each other in the later training stage. Eq. (5)) proxies for each sample during back-propagation, and we use a small batch size As future work, we will focus more on addressing such datasets with huge inter-class variance.

Weaknesses: Although the evaluation datasets are sound and reliable, I am concerned with overclaiming. The paper claims at multiple occasions to achieve state-of-the-art results (lines 14, 63, 273) or even substantially outperform them (line 199). These claims seem reasonable given the reported results, but I think that they are actually not valid: 1. The baselines that are compared against are mostly other dense matching networks (Sparse-/NCNet in Figures 3, 6, and 7) or learned local features (D2Net, R2D2, SuperPoint in Figures 5 and 7). The evaluation does not include deep neural networks for sparse feature matching, which, similarly to DRC-Net, leverage information from both images, and are already referenced in the related work (line 76).

Weaknesses: (Validation of novelty) Leveraging texture and shape consistency for online refinement has been used in prior literature, e.g., [3, 4] in a different context and application. I believe the most significant contribution is the way texture and shape invariance is applied, i.e., swapping pose and texture across time to enforce consistency. However, its novelty is not empirically validated at all, i.e., why is this design choice reasonable and better than the existing idea. For instance, they argue that their idea can address blurry texture and low res shape while the results look blurred and low res anyway. Without rigorous ablation study in isolation, it is not clear how the core novelty makes an impact.

Weaknesses: The rebuttal and discussion clarified my concerns about [1,2] (although I would highly encourage that these works be citied for a more complete related works section). However, I remain unconvinced by the novelty of the approach -- the fact that transformer based models work better compared to simple VAE based models is not surprising to the general NeurIPS audience. However, I do agree that from the point of view of stroke based generative models the work is novel and makes a good contribution to this specific field. Novelty wrt to [1] is not clear -- both methods use a transformer based architecture to model long-range dependencies in strokes. The advantage of an autoregressive structure along with transformers is not clear as transformers contain self-attention layers to capture long range dependencies.

目前在電腦視覺的領域上，具備深度的卷積神經網路已經在多種任務上取得很大的進步。近年在 Convolution block 的設計上，有發展出一系列相當有潛力的研究，主要是探討將 Attention 機制應用於 Channel 維度的作法。 其中相當具代表性的研究是 SENet，後續也有一些研究嘗試利用更複雜的機制捕捉 Channel-wise dependencies ，或結合 Spatial…

SenseTime is focused on clinical needs to empower the entire workflow of "diagnosis-treatment-recovery". SenseTime is focused on clinical needs to empower the entire workflow of "diagnosis-treatment-recovery".

In the last few years, we have witnessed a renewed and fast-growing interest in continual learning with deep neural networks with the shared objective of making current AI systems more adaptive, efficient and autonomous. However, despite the significant and undoubted progress of the field in addressing the issue of catastrophic forgetting, benchmarking different continual learning approaches is a difficult task by itself. In fact, given the proliferation of different settings, training and evaluation protocols, metrics and nomenclature, it is often tricky to properly characterize a continual learning algorithm, relate it to other solutions and gauge its real-world applicability. The first Continual Learning in Computer Vision challenge held at CVPR in 2020 has been one of the first opportunities to evaluate different continual learning algorithms on a common hardware with a large set of shared evaluation metrics and 3 different settings based on the realistic CORe50 video benchmark. In this paper, we report the main results of the competition, which counted more than 79 teams registered, 11 finalists and 2300$ in prizes. We also summarize the winning approaches, current challenges and future research directions.