HackerNews Summary (2023-06-12 09PM)

Source: Xerox scanners/photocopiers randomly alter numbers in scanned documents

Summary
Xerox WorkCentre and ColorQube scan copiers have an eight-year-old bug that randomly alters numbers in scanned documents, even if OCR is turned off. This can lead to dangerous scenarios, such as incorrect invoices and construction plans, and incorrect metering of medicine, potentially putting lives at risk. The bug affects hundreds of thousands of machines globally.

Facts

🖨️ Xerox WorkCentre and ColorQube scan copiers have an eight-year-old bug that randomly alters numbers in scanned documents.

🔍 The bug affects hundreds of thousands of machines globally, and has potentially caused incorrect invoices, construction plans, and metering of medicine.

📖 Despite having OCR turned off, patches of pixel data are randomly replaced, leading to incorrect information being disseminated through scanned documents.

🚨 The errors can pose a danger to life, particularly in the case of incorrect construction plans or metering of medicine

👨‍💼 The last 8 years of PDF scans of affected devices may not only contain errors, but may also have no legal value, even if proven true.

---

Source: Meta to Buy Almost 7 Million Carbon Credits from Aspiration

Summary
Facebook’s parent company Meta has agreed to purchase 6.75 million carbon removal credits from Aspiration Partners to scale up nature-based climate solutions. The credits will be based on various ecosystem restoration projects, such as agroforestry and sustainable agriculture practices.

Fact

🌳 Meta has committed to achieving net zero emissions across its entire value chain by 2030 through carbon removal initiatives that align with Science-Based Targets initiative standards.

🌍 Aspiration aims to provide high-quality nature-based carbon credits that focus on environmental and social benefits. Partnering with top-tier project developers and independent verifiers, the sustainability firm ensures transparent and verifiable carbon removal credits.

🌿 For Aspiration CEO, Olivia Albrecht, corporate carbon commitments amplify climate goals and make a significant impact.

---

Source: Oyster: Towards Unsupervised Object Detection from LiDAR Point Clouds – Waabi

Summary
This paper proposes OYSTER, an unsupervised object detection method from 3D point clouds in self-driving scenes. It uses point clustering, temporal consistency, CNNs for long-range extension, and self-supervision. It does not require constraints on data collection and is zero-shot without supervised fine-tuning. The proposed planning-centric perception metric outperforms unsupervised baselines on PandaSet and Argoverse 2 Sensor datasets.

Fact

🔍 OYSTER is an unsupervised object detection method from 3D point clouds in self-driving scenarios.

🧮 It exploits point clustering, temporal consistency, translation equivariance, and self-supervision for improving on its own.

🚗 It does not impose constraints on data collection and is zero-shot without supervised fine-tuning.

📈 OYSTER outperforms unsupervised baselines on PandaSet and Argoverse 2 Sensor datasets.

📊 A planning-centric perception metric based on distance-to-collision is proposed to better measure model performance in self-driving scenarios.

---

Source: SnapFusion

Summary
Researchers have developed an approach that allows text-to-image diffusion models to run on mobile devices in under two seconds. By introducing an efficient UNet network architecture and improving step distillation, the team achieved better FID and CLIP scores than previous models with significantly fewer denoising steps.

Fact

💻 The new approach allows text-to-image diffusion models to run on mobile devices in less than two seconds.

📱 The network architecture was made more efficient, reducing computation for the image decoder via data distillation.

🎨 The new step distillation method introduces training strategies and regularization from classifier-free guidance, achieving better FID and CLIP scores than previous models.

👨‍💻 The approach democratizes content creation by making powerful text-to-image diffusion models more accessible to users.

---