【Paper material】

---

使用

• employ …… to （use……to）（用什么）
• utilize （用）
• leverage （利用）
• yielding （产生）
• super-impose （叠加） v，例如：we super-impose the channels from the different keypoints

灵感来自于

• In this paper, we propose an architecture that distills this insight into a simple connectivity pattern （灵感来自于）
• Overall, SSD shares insights with many concurrent works （灵感）
• by building in prior knowledge （先验知识）
• These successes spurred a new line of research that focused on finding higher performing convolutional neural networks.（刺激了新的研究方向）
• Following that intuition （跟着直觉）
• The design of our search space took much inspiration from LSTM, and Neural Architecture Search Cell. （从XXX中取灵感）

提出，猜测，发现

• introducing redundant degrees of freedom and increasing the risk of overfitting （引入）
• propose （提出）
• present （提出）
• derive a weight initialization scheme （引入了XXX方案）
• We emphasize that （强调）
• In light of these observations, we can conclude that （得出结论）
• We conjecture that （推测、猜想）
• We surmise that （我们推测）
• We empirically found （经验发现）
• both being derived under the assumption that （基于……假设推导出来的）

被称为

• (Region Proposal Network, RPN for short)（简称）
• We propose a deep convolutional neural network architecture codenamed Inception that achieves the new state of the art for classification and detection in the (ILSVRC14). （代号）
• XXX，dubbed Xception （XXX 被称为）
• The above operation can be recast as （上面的操作等价于（重定义为））

缓解问题

• alleviate the vanishing-gradient problem （缓解了什么问题）
• mitigates this issue to an extent （在某种程度上缓解了这种问题）
• propose a new loss function that eliminates this barrier （消除障碍）
• surpassed the 100-layer barrier

效果提升

• substantial enhancement （大幅度增强）
• improve significantly （显著提高）
• significantly outperform （明显优于）
• and can further exceed its accuracy by a substantial margin.（大大超过其准确性）
• further improve （进一步提升性能）
• improve the results only marginally （少量的提升）
• shows marginally better results （显示稍微好一点的结果）
• shows small（large） gains in classification performance on the ImageNet dataset （效果好一点点 / 好很多）
• Our results compare favorably with ResNet （结果比ResNet更好一点）
• DenseNets have several compelling advantages （引人入胜的优点）
• we mark all results that outperform the existing state-of-the-art in boldface and the overall best result in blue.（傲世群雄）
• we secured second place （我们获得了第二名）
• achieves both higher mAP quality and lower latency than MobileNet （更高的精度，更快的速度）
• In order to enhance the capability of the network

效果不好

• just on par with the （与什么同水平）
• is marginally worse than （差一点点）
• any adverse effect （副作用）
• This problem becomes even more pronounced once pooling units are added to the mix:（问题更严重）
  but have trailed the accuracy of two-stage detectors thus far.（one stage 比 two stage 精度低）
• but have trailed the accuracy of two-stage detectors thus far.（已经落后）
• have trailed the accuracy （落后的精度）

速度和精度的权衡

• lightweight models （轻量级模型）
• with negligible computation overhead. **（可忽略不计的计算开销）**with negligible computation overhead. （可忽略不计的计算开销）
• it’s time consuming （耗时的）
• memory and time consuming.（内存和时间消耗）
• can be prohibitively expensive （极其昂贵）
• that efficiently trade off between latency and accuracy.（有效的权衡延迟和准确性）
• This factorization has the effect of drastically reducing computation and model size.（大幅减少计算和模型大小）
• Residual connections are clearly essential in helping with convergence, both in terms of speed and final classification performance. （在速度和精度上，帮助收敛）
• is marginally slower （稍慢一些）
• strike an excellent trade-off between representation capability and computational cost. （在表示能力和计算成本之间取得很好的平衡。）
• Real world tasks often aim at obtaining best accuracy under a limited computational budget,（在有限的资源上取得最好的精度）
• Table 1 shows the theoretical computational cost. Though the actual time cost might be influenced by other factors, e.g. GPU bandwidth and coding quality, the computational cost shows the speed upper bound.（理论的计算代价）

连接词

• Consequently （因此）
• Crucially （至关重要的）
• In addition （另外）
• Aside from being capable of representing higher-level semantics,（除此之外）
• Unless otherwise specified （除了有另外的说明）
• Analogous to （类似于）
• For instance （例如）
• One principal reason is that （一个主要的原因是）
• In this light （就此而论）
• In a more succinct reformulation （用更简洁的语言重新表达一次）
• Specifically （具体而言）
• More concretely （具体而言）
• in a nutshell （简而言之）
• in line with （与XX一致）
• the discrepancy between （XXX之间的差别）
• judiciously reducing dimension wherever the computational requirements would increase too much otherwise （明智的）
• The output of a network processing such data should be equivariant w.r.t. the orientation of its input.（关于）—— with respect to
• The results presented in the figure reveal that DenseNets perform on par with the state-of-the-art ResNets, whilst requiring significantly fewer parameters and computation to achieve comparable performance.（同时）
• Fast and Faster R-CNN [11, 29] opt to not use featurized image pyramids under default settings.
  （opt to 选择 opt to not 选择不）
• prone to. （倾向于）
• reliant on （依赖于）
• To the best of our knowledge （据我所知）

值得注意的是

• It is noteworthy that （值得注意的是）
• It is worth noting that （值得注意的是）
• Note that, even when （注意，即使是）
• that it is nontrivial for them to achieve good accuracy.（重要的）
• is sorely needed. （是非常必要的）
• Most remarkably （最值得注意的是）

图

• Thin curves （细线）
• bold curves （粗线）
• Figure 1 illustrates the pipeline （图1展示了）
• Figure 1 illustrates this layout schematically. （布局示意图）
• see Figure 7 for schematic （原理图）
• Viewing digitally with zoom is recommended. （推荐放大看）
• schematic view of XXX （XXX的示意图，网络结构）

---

1. COCO，80k train images and a 35k subset of val images (trainval35k), and report ablations on
   the remaining 5k subset of val images (minival).
   ILSVRC2013，train（395,918），val（20,121），test（40,152），200classes
   We comprehensively evaluate our method on the Caltech pedestrian detection benchmark [9] and PASCAL VOC 2007 object detection benchmark.（综合评估）
   evaluation metrics （评价指标）
   the canonical ImageNet pre-training size （典范的，标准的形式）
   the canonical form of an Inception module （典范的，标准的形式）
   benchmark datasets （公共数据集）
2. Each image was sampled randomly(there are not curated（拉丁文）). （没有什么策划）
3. coarser granularity （粗粒度）
   fine granularity （细粒度）
   class-specific and class-agnostic （类别）
   Symmetry and asymmetry （对称性和不对称性）
4. element-wise addition （元素方式的加法）
5. This product may also impede information propagation and hamper the training procedure as witnessed in the following experiments.（妨碍，阻碍）
6. tracking the maximum of the sharpness of the images （聚焦最优值）
   any change of focal distance （聚焦）
   in-focus image （聚焦）
   As the sharpness is a function of the working distance, autofocus may be considered as a prob- lem of optimization: the search for the peak of sharpness along the optical axis.（聚焦）
7. hand-engineered features （手工设计的特征）
8. aforementioned （上述提及的）
9. the rest of the approach is similar to vanilla ResNet （原始的）
10. which is beyond the scope of this paper （超过了这篇文章的讨论范围）
11. The single-stage detector usually targets on a sweet-spot of very fast speed and reasonably good accuracy《light head》（目标，目的）
12. Since the weights to be learned serve as expansion coefficients for the steerable function space, common weight initialization schemes need to be adapted. （改变权重初始化方案）
13. The main hallmark of this architecture is XXX （n. 特点；品质证明）
14. We hope our approach can facilitate future research on training very tiny CNNs for cutting-edge applications. （前沿的应用）
15. Before 2012, specialized solutions were required for each specific application domain. Since then, deep convolutional neural networks have become mainstream to address these tasks.（成为主流）
16. We also discuss drawbacks of the existed methods in scenarios of driver distraction detection.
17. provide a viable search strategy （提供了一个可行的搜索策略）
18. Image classification networks provide generic image features that may be transferred to other computer vision problems.参考文献：A deep convolutional activation feature for generic visual recognition （分类的重要性）
19. The best single run （跑得最好的模型）
20. and makes it notoriously hard to train models with saturating nonlinearities.（出了名的困难）
21. off-the-shelf CNN models （现有的模型）
22. concurrently、simultaneously （同时地）
23. as far as we know