尼欧 ᯤ

To no avail but try.

[Ray Wang @wangray]

技术再好，讲不清楚就等于没有

Patrick Winston 教授这个讲座我看了不下三遍，做产品、做内容、见客户，说话是最被低估的杠杆

他在 MIT 将这个课超过40年，这是最后一次录像，他于2019年去世了

他的核心观点：演讲能力不是天赋，是可以训练的手艺

一小时看完，值得

[JUMPERZ @jumperz]

TL;DR

[Tz @Tz_2022]

一张图看懂 2028 全球智能危机全部内容

[尼欧 ᯤ @________Neo]

@Cydiar404 Mark

[尼欧 ᯤ @________Neo]

@mehu_wa @HuoQubot

[Growth Labs @growthhub_]

Bookmark it.

[尼欧 ᯤ @________Neo]

@sehui2534690907 @HuoQubot

[Alex Jones @RealAlexJones]

DESTROYED BY DESIGN:

Los Angeles Fires Are Part Of A Larger Globalist Plot To Wage Economic Warfare & Deindustrialize The Untied States Before Triggering Total Collapse

» WATCH/SHARE THE LIVE X STREAM HERE:
https://t.co/2jcVYLsnzw

[尼欧 ᯤ @________Neo]

the zero-shot accuracy matters more

[Aakash Kumar Nain @A_K_Nain]

Zero-shot monocular depth estimation in real-time is one of the most challenging problems in vision. Applications like novel view synthesis from a single image require a strong depth estimation model. The paper DepthPro from Apple presents another model that can perform zero-shot depth estimation on high-resolution images with low latency. Here is a quick summary in case you are interested:

Desired characteristics of a depth estimation model
- It should not be restricted to a single domain and should produce zero-shot depth estimation on any image.

- The model should produce metric depth maps in a zero-shot regime to accurately reproduce object shapes, scene layouts, and absolute scales.

- The model should produce metric depth maps with absolute scale even if no camera intrinsics (such as focal length) are provided with the image. It helps enable novel view synthesis from an arbitrary image.

- The model can operate on high-resolution images and should be able to produce high-quality depth maps even for complex objects like hair, fur, etc.

- It should run with extremely low latency to support applications like view synthesis on demand.

The DepthPro model ticks all the things listed above.

Network
- Uses pretrained Vision Transformers (ViTs) as encoders.

- The key idea is to apply plain ViT encoders on patches extracted at multiple scales and fuse the patch predictions into a single high-resolution dense prediction in an end-to-end trainable mode.

- Two encoders in total: Patch encoder and image encoder.

- The patch encoder is applied on patches extracted at multiple scales, allowing the model to learn scale-invariant representations as weights are shared across scales.

- The image encoder is applied to the whole image, downsampled from the original resolution to the resolution accepted by the base encoder (384 x 384 in this case). The image encoder anchors the patch predictions in a global context.

- The network operates at a fixed resolution of 1536 × 1536 chosen as a multiple of the ViT’s 384×384.

- The original image with a resolution of 1536 x 1536 is processed with two more downsampled resolutions of 784 x 784, and 384 x 384. For each resolution, the image is divided into patches of size 384x384.  At each scale, the patches are fed into the patch encoder that produces a feature tensor at a resolution of 24 × 24 per input patch. Intermediate features for the finest scale are also extracted to capture fine-grained details.

- The feature patches are then merged into maps and are fed into a DPT-like decoder module.

Training Objectives
- The network f predicts a canonical inverse depth image C = f (I) for an input image I. The dense metric depth map is then obtained as D_m = f_px / wC, where f_px is the focal length and w is the width.

- All objectives use canonical inverse depth as it prioritizes areas close to the camera over farther areas or the whole scene. This supports visual quality in applications such as novel view synthesis.

- For all metric datasets, the authors compute mean absolute error (MAE) per pixel i and discard pixels with an error in the top 20% per image for real-world (not for synthetic) datasets.

- For all non-metric datasets (i.e., those without reliable camera intrinsics or inconsistent scale), the authors normalize predictions and ground truth via the mean absolute deviation from the median before applying a loss. They also define a multi-scale derivative loss over M scales. The ∇∗ indicate a spatial derivative operator ∗, such as Scharr or Laplace (L), and p is the error norm. The scales j are computed by blurring and downsampling the inverse depth maps by a factor of 2 per scale.

I did not find this in the code, but I have asked the authors to give some implementation details. Let us see how it goes.

Training Curriculum
The authors note down three important observations before launching a training run:

- Training on a large mix of real-world and synthetic datasets improves generalization.

- Real-world datasets are messy, and many times the labels are noisy (missing areas, mismatched depth, or false measurements on object boundaries.)

- Predictions get sharper throughout the training.

Based on these observations, the authors design a two-stage training curriculum. In the first stage, they aim to learn robust features that allow the network to generalize across domains, and train the model on a mix of all labeled training sets. They use MAE for the metric datasets and the normalized MAE for non-metric datasets. They also apply scale-and-shift-invariant loss on gradients, but only to synthetic datasets.

The second stage of training is designed to sharpen boundaries and reveal fine details in the predicted depth maps. Given that the synthetic dataset provides high-quality pixel-accurate ground truth, the authors use it to minimize the effect of inaccurate ground truth. MAE is supplemented with Mean Absolute Gradient Error (MAGE), Mean Absolute Laplace Error (MALE), and the Mean Squared Gradient Error (MSGE).

Evaluation metrics for sharp boundaries.
- Common benchmarks for monocular depth prediction rarely take boundary sharpness into account. This may be attributed in part to the lack of diverse and realistic datasets with precise pixel-accurate ground-truth depth.

- The authors aim to leverage existing high-quality annotations for matting, saliency, or segmentation as ground truth for depth boundaries.

- They treat annotations for these tasks as binary maps, which define a foreground/background relationship between an object and its environment. To ensure that the relationship holds, they only consider pixels around edges in the binary map.

- The hypothesis is that if the depth of pixel i and pixel j differs by more than t%, it suggests the presence of an occluding contour between those pixels. Occluding contour c_d(i, j), and overall precision and recall for the neighboring pixels are then calculated as shown below:

- The authors report the weighted F1 score with thresholds that range linearly from tmin = 5 to tmax = 25, with stronger weights towards high threshold values.

- The beauty of this method is that it doesn't require any manual edge annotations. It only needs pixel-wise ground truth that can be easily obtained for synthetic datasets.

- The authors also apply non-maximum suppression to values of c ˆd within the valid bounds of c ˆd(i, j) connected components.

- Occluding contours from binary maps are similarly obtained

Focal Length Estimation
- Trained the network to predict focal length to handle cases where the EXIF metadata is either missing or is noisy for the input image.

- The focal estimation head is a small convolutional network that ingests frozen features from the depth estimation network and task-specific features from a separate ViT image encoder to predict the horizontal angular field of view.

- This head is trained after the training for depth prediction is complete.

Results

Limitations
The model is limited in dealing with translucent surfaces and volumetric scattering, where the definition of single-pixel depth is ill-posed and ambiguous.

[尼欧 ᯤ @________Neo]

保护老腰

[Brandon Scott @NaturalMVMT]

I spent 6 years at the chiropractor for low back & upper back/neck pain, with little luck.

This is the method I used to rebuild my low back & finally get the relief I was looking for (Part 2)...
🧵🧵🧵

[章工 @435hz]

保护意志力最关键的是不能丧失对生活的希望，否则再多的技巧也没用。
中年无子女、工作遇到困境、夫妻关系矛盾无解、财富重大损失、慢性病……都容易让人失去希望，只想行尸走肉活一辈子。甚至外遇对有一些人是生活的一剂春药。
每个人都要找到属于自己的保持斗志的方法，特别是那些已经丧气的人。

[硅谷王川 Chuan @Svwang1]

同一件事情给人带来的愉悦感的深浅程度，取决于人体内多巴胺受体 (dopamine receptor) 的数目. 多巴胺受体，是细胞表面用来接受和传输多巴胺信号的蛋白质。

多巴胺受体越多，多巴胺的传输就越流畅，人就自然感觉越舒服。

适度运动可以对大脑神经元内多巴胺受体的生成产生立竿见影的效果。而精神压力导致体内皮质醇分泌，则会抑制多巴胺受体的生成，这样人可以集中精力应对面前的麻烦，而不是浪费资源寻找多巴胺的快乐。

小孩子之所以日常很快乐，一是经常蹦蹦跳跳，运动量比成年人大 ；二是精神压力相对成年人更小，或者根本不理解一些麻烦可能带来的严重后果。

所以，保持日常极为愉悦的精神状态，需要在两方面努力：

第一，在不受伤的前提下，多运动，多样化的运动！

第二，足够冷酷，对自己足够知识诚实 (intellectually honest) 而不是掩耳盗铃，未雨绸缪，提前把任何潜在的麻烦和导致精神压力的事情控制在可以承受的范围内。

[章工 @435hz]

提高自信心指数比较有效的事：

1.把一件曾经不擅长的事做擅长
2.把一件新接触的事做擅长
3.把一件已经擅长的事做得更擅长

顺序也是这个顺序，自信的剂量主要取决于你在这件事里突破了多少自我。

[章工 @435hz]

人一辈子就一件事，找到属于你的赛道，做到极致。
有人擅长商业，有人擅长学术，有人擅长业务……总有适合自己的。

有些人很幸运，年轻时就找到了；大部份人找了一辈子也没找到，虚度了人生，叫一辈子没离开人生online新手村，糊里糊涂就销号了。

[游戏人间 @Cryptoeth10w]

永远不要觉得尴尬和丢脸

[章工 @435hz]

怎么判断网上的东西值不值得关注，该不该花钱?
其实很简单，你就看这些东西能不能解决你的实际问题，这个你比任何人都清楚。

[Odyssey @OdysseysEth]

无知是没有正确的信息和算法，而愚蠢是有了错误的信息和算法。

[章工 @435hz]

有些问题，沉默就是答案，躲闪就是答案，没有主动回应就是答案。

[Tristan @wsygc]

真正要搞起一个互联网产品，要做这么多事儿呀，开发只是其中一环...