All Articles

A flat index of all articles in this wiki. Add a new line here whenever a new article is created. For a human-readable entrypoint with reading paths and topic groupings, see Home.

Articles

• Attention Is All You Need — The original Transformer: multi-head self-attention encoder-decoder, no recurrence or convolution (Vaswani et al., NeurIPS 2017)
• DETR — End-to-end object detection with transformers (Carion et al., 2020)
• YOLOv8 — Anchor-free real-time detector from Ultralytics (2023), baseline for YOLOv10
• YOLOv10 — Real-time end-to-end object detection with NMS-free training (Wang et al., 2024)
• Task-Aligned-Assignment — Dynamic label assignment via the metric $t=s^{\alpha }{u}^{\beta }$ (TOOD, Feng et al., 2021); used by YOLOv8/v10
• Implementing Research Papers — Practical guide/workflow for turning a paper into working code
• Publishing the Wiki — How this vault is published to wiki.krystex.cc and how to push updates
• Spanisch Reflexive Verben — Reflexive Verben im Spanischen: Pronomen, Stellung, Verwendungstypen (deutschsprachig)
• Integrals — Tutorial: what integrals compute, Riemann sums, the Fundamental Theorem of Calculus, and substitution/parts
• SAM — Segment Anything: promptable segmentation foundation model + the 1.1B-mask SA-1B dataset (Kirillov et al., 2023)
• SAM2 — Segment Anything 2: promptable video segmentation with streaming memory + the SA-V dataset (Ravi et al., 2024)
• DETR Family Comparison — Side-by-side comparison of all eight DETR-line detectors: year, COCO AP, params, FLOPs, throughput, training cost
• MobileSAM — Lightweight SAM via decoupled distillation of the image encoder into a 5M-parameter ViT-Tiny (Zhang et al., 2023)
• Deformable DETR — Sparse multi-scale deformable attention for DETR: 10× faster convergence and much better small-object AP (Zhu et al., ICLR 2021)
• Conditional DETR — Conditional spatial queries for decoder cross-attention: 6.7–10× faster DETR training with global dense attention kept (Meng et al., ICCV 2021)
• DAB-DETR — 4D anchor boxes as decoder queries, with width/height-modulated attention and layer-wise anchor update (Liu et al., ICLR 2022)
• DN-DETR — Denoising training: noised GT boxes as auxiliary queries to stabilise Hungarian matching (Li et al., CVPR 2022)
• DINO — Contrastive denoising, mixed query selection and look forward twice; first DETR to top COCO (Zhang et al., ICLR 2023). Not Meta’s self-supervised DINO
• RT-DETR — Real-time DETR: efficient hybrid encoder + uncertainty-minimal query selection, beating YOLOs without NMS (Zhao et al., CVPR 2024)
• LW-DETR — Light-weight DETR: plain ViT encoder with interleaved window/global attention, 3-layer decoder and Objects365 pre-training; 42.6–58.3 mAP at 500–52 FPS (Chen et al., 2024)
• GPUs for Local ML — Cost, VRAM, FLOPs, and bandwidth comparison of the M4 Pro, RTX 3060/3090/5090, and DGX Spark
• SAM3 — Segment Anything with Concepts: noun-phrase and exemplar prompts segment every matching instance, via a DETR detector with a presence head plus a SAM 2 tracker and the AI-verified SA-Co data engine (Meta, 2025)
• TinyViT — 5M/11M/21M hybrid conv-transformer backbones plus fast pretraining distillation from stored sparse teacher logits; the backbone MobileSAM distils SAM into (Wu et al., ECCV 2022)
• Your ViT is Secretly an Image Segmentation Model — EoMT: a plain pre-trained ViT plus query tokens does panoptic segmentation with no adapter, pixel decoder or Transformer decoder, ~4× faster than ViT-Adapter + Mask2Former (Kerssies et al., CVPR 2025)
• PerSAM — Personalize SAM from one image+mask: training-free location prior, attention guidance and semantic prompting, plus a 2-parameter scale-aware fine-tuning (Zhang et al., ICLR 2024)
• HSNet — Hypercorrelation Squeeze: few-shot segmentation by squeezing a pyramid of multi-level 4D correlation tensors with efficient center-pivot 4D convolutions (Min et al., ICCV 2021)
• RefT — Reference Twice: an RPN-free Mask2Former baseline for few-shot instance segmentation that references the support set at both feature and object-query level, plus a class-enhanced BKD loss for the incremental setting (Han et al., arXiv 2024)
• Few-Shot Segmentation Comparison — Taxonomy and side-by-side comparison of HSNet, RefT, PerSAM and SAM 3: task output, shot type, training regime, why the benchmark numbers do not compare, and how SAM turned few-shot segmentation into a naming problem
• YOLACT — Real-time instance segmentation via prototype masks + per-instance linear coefficients, assembled by one matrix multiply (Bolya et al., ICCV 2019); the mask head YOLOv8-seg reuses
• Mask R-CNN — RoIAlign + a per-class FCN mask branch added to Faster R-CNN; the two-stage instance segmentation baseline (He et al., ICCV 2017)
• SOLOv2 — Box-free instance segmentation via dynamic per-grid-cell convolution kernels and a unified mask feature map, plus Matrix NMS (Wang et al., NeurIPS 2020)
• Segmentation Head Comparison — Five instance/panoptic segmentation mask heads (Mask R-CNN, YOLACT, SOLOv2, DETR, Mask2Former) compared side by side, with original architecture diagrams for each
• SigLIP — Replaces CLIP’s softmax contrastive loss with a simpler pairwise sigmoid loss, enabling efficient, memory-light language-image pre-training at small-to-moderate batch sizes (Zhai et al., 2023)
• SigLIP2 — Multilingual SigLIP successor combining sigmoid loss, LocCa captioning and SILC/TIPS self-distillation into one recipe, plus NaFlex native-resolution support and stronger localization/dense features (Tschannen et al., 2025)
• How to train your ViT — Large-scale AugReg study: augmentation and regularization substitute for ~10x more pre-training data, matching JFT-300M results using only public ImageNet-21k (Steiner et al., 2021)
• Image Captioners Are Scalable Vision Learners Too — Compute-matched study showing plain image-captioning pretraining (Cap/CapPa) rivals or beats CLIP-style contrastive pretraining, especially on captioning, VQA and fine-grained transfer (Tschannen et al., 2023)
• Mask2Former — Universal mask-classification architecture with a masked-attention Transformer decoder that beats specialized SOTA on panoptic, instance and semantic segmentation with one architecture (Cheng et al., CVPR 2022)
• PaliGemma — A 3B VLM combining SigLIP-So400m and Gemma-2B via a linear connector and prefix-LM attention, trained as a versatile transfer base across captioning, VQA, detection/segmentation-as-text and video (Beyer et al., 2024)
• PANet — Few-shot segmentation via masked-average-pooled class prototypes and cosine-distance matching, with prototype alignment regularization (PAR) that segments support images back from predicted query masks (Wang et al., ICCV 2019)
• Matcher — Training-free one-shot segmentation pairing DINOv2 correspondence with SAM via bidirectional matching and instance-level matching, no fine-tuning (Liu et al., 2023)
• DE-ViT — Few-shot/one-shot object detector using frozen DINOv2 features, region propagation and prototype-subspace projection, with no test-time finetuning (Zhang et al., Rutgers, CoRL 2024)
• x86 Assembly Fundamentals — Registers, flags, syntax and addressing modes of x86/x86-64, framed against ARM for readers coming from that background
• x86 Instructions and the Stack — Core instructions, conditional jumps, the stack, and calling conventions (cdecl, stdcall, System V, Microsoft x64)
• Reading Compiled Code Patterns — How if/loops/switch/structs/arrays/vtables look once compiled to x86, for reading disassembly back to source-level intent
• x86 Reverse Engineering for Games — Disassemblers/debuggers, finding functions of interest, patching techniques, and a worked crackme walkthrough
• Cheat Engine and Memory Scanning — Memory scanning workflow, pointer chains surviving ASLR/restarts, and a trainer-style memory-edit walkthrough
• PEM — Prototype-based Efficient MaskFormer: one-pixel-per-query prototype cross-attention and a convolutional CSM/deformable-conv pixel decoder make Mask2Former-style semantic+panoptic segmentation 2–3× faster
• EMA — Exponential moving average of weights: the update rule, effective window, warm-up ramps, BN handling, a minimal PyTorch implementation, and its use as a momentum teacher (Mean Teacher, MoCo, BYOL, DINO)
• YOLOE — Open-vocabulary detection + segmentation in one YOLO: re-parameterizable text alignment, SAVPE visual prompts, lazy retrieval for prompt-free (Wang et al., Tsinghua, 2025)
• Tesla Model 3 — Used-buyer guide for Germany: model-year changes, LFP vs NCA/NMC packs, charging, warranty, TÜV weak spots, battery health checks and a purchase checklist
• Polestar 2 — Used-buyer guide for Germany: variants and MY24 RWD facelift, battery and real-world range, recalls, Öhlins/TCAM/12 V cost traps, TÜV record, service network
• Tesla Model 3 vs Polestar 2 — Side-by-side of the two used EVs: efficiency, charging, reliability (TÜV), running costs, and a decision guide
• Hyundai Ioniq 5 — Used-buyer guide for Germany: 800 V E-GMP platform, variants through the 2024 facelift, ICCU recall, charging, TÜV record, prices
• Kia EV6 — Used-buyer guide for Germany: Ioniq 5 twin on E-GMP, GT and facelift, ICCU/zombie-mode issues, 7-year warranty, prices
• ArcFace — additive angular margin softmax for face recognition: normalised embeddings, $\cos (\theta +m)$ target logit, sub-center variant for noisy web data, and BN-prior model inversion