papershelf
Papershelf
Papers I keep coming back to — each with a one-line takeaway so you can scan the whole shelf in a minute.
12 papers · 2004–2023 · 5 areas
- arXiv:2308.04079 SIGGRAPH · 2023
3D Gaussian Splatting for Real-Time Radiance Field Rendering
Replaces ray marching with rasterized 3D Gaussians — radiance fields at real-time frame rates.
- arXiv:2304.02643 ICCV · 2023
Segment Anything
SAM — a promptable foundation model that segments anything, zero-shot, in any image.
- arXiv:2201.05989 SIGGRAPH · 2022
Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
A multiresolution hash encoding that cuts NeRF training from hours to seconds.
- arXiv:2010.11929 ICLR · 2021
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
ViT — split an image into patches, feed it to a plain transformer, and beat CNNs at scale.
- arXiv:2003.08934 ECCV · 2020
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
Encodes an entire scene in an MLP and renders photoreal novel views by ray marching through it.
- arXiv:1712.05877 CVPR · 2018
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference
The int8 quantization-aware training recipe that puts float-quality models on integer-only hardware.
- arXiv:1706.03762 NeurIPS · 2017
Attention Is All You Need
The transformer — self-attention replaces recurrence and convolution as the dominant architecture.
- arXiv:1704.04861 arXiv · 2017
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision
Depthwise separable convolutions — ~9x fewer FLOPs with modest accuracy loss, built for phones.
- arXiv:1512.03385 CVPR · 2016
Deep Residual Learning for Image Recognition
ResNet — skip connections let gradients flow through 100+ layer networks that actually train.
- arXiv:1506.02640 CVPR · 2016
You Only Look Once: Unified, Real-Time Object Detection
YOLO — detection as a single regression pass, making real-time perception practical on modest hardware.
- NeurIPS 2012 NeurIPS · 2012
ImageNet Classification with Deep Convolutional Neural Networks
AlexNet — GPUs + ReLU + dropout halved the ImageNet error rate and started the deep learning era.
- IJCV 2004 IJCV · 2004
Distinctive Image Features from Scale-Invariant Keypoints
SIFT — keypoints and descriptors invariant to scale and rotation, the backbone of two decades of matching.