Computer Vision

Descriptors and correspondence

  • SIFT - Extracts local descriptors invariant to scale and rotation for matching objects or scenes across images; robustness to viewpoint and illumination changes is empirical, not unlimited invariance. Lowe 2004.
  • Histograms of oriented gradients - Pools local gradient orientations with overlapping block normalization. The human-detection study isolates how spatial binning, orientation resolution, and contrast normalization affect recognition. Dalal and Triggs 2005.
  • SuperPoint - Jointly predicts interest points and descriptors from full images; Homographic Adaptation supplies self-supervision for repeatable matching across transformed views. DeTone et al. 2018.
  • SuperGlue - Uses attention and a graph neural network to predict matching costs, then differentiable optimal transport to assign local features and reject unmatched points. Sarlin et al. 2020.
  • LoFTR - Matches images without a separate keypoint detector: cross-image attention establishes coarse correspondences, then refines selected matches at higher resolution. Sun et al. 2021.

Classification backbones

  • Cresceptron - Grows a hierarchy of neurons and connections as new concepts appear. Uses hierarchical tolerance for recognition and traces response paths for segmentation. Weng et al. 1992.
  • AlexNet - Trains a deep convolutional classifier on ImageNet using GPU computation, non-saturating neurons, and regularization; a reference for large-scale supervised visual learning. Krizhevsky et al. 2012.
  • VGG - Studies depth using stacks of small 3×3 convolution filters; provides classification networks whose learned representations transfer to other image datasets. Simonyan and Zisserman 2015.
  • ResNet - Learns residual functions relative to layer inputs, easing optimization of deep networks. The paper evaluates recognition and transfers the backbone to detection. He et al. 2016.
  • DenseNet - Feeds each layer the feature maps of preceding layers within densely connected blocks, promoting feature reuse and short paths for information and gradients. Huang et al. 2017.
  • EfficientNet - Scales network depth, width, and input resolution together using a compound coefficient; combines this scaling rule with a searched convolutional baseline. Tan and Le 2019.
  • Vision Transformer - Applies a Transformer to image-patch sequences. The study emphasizes large-scale pretraining and transfer rather than assuming attention needs less data than convolutions. Dosovitskiy et al. 2021.
    • DeiT - Trains image Transformers on ImageNet without external training data; an attention-based distillation token transfers knowledge from a teacher, including a convolutional teacher. Touvron et al. 2021.
    • Swin Transformer - Builds hierarchical image features using local attention windows shifted between layers. Cross-window connections retain spatial interaction while limiting attention cost. Liu et al. 2021.
  • ConvNeXt - Revisits convolutional design by progressively modifying a ResNet toward choices used in visual Transformers; supplies convolutional backbones for classification and dense prediction. Liu et al. 2022.

Learning reusable visual features

  • Self-supervised representations - See Feature Extraction for contrastive learning, BYOL, DINO and masked autoencoders, with their direct analyses and follow-ups.
  • Vision-language models - CLIP supports language-defined recognition; Flamingo and LLaVA generate text from visual context. See Multi-View Learning for alignment and generation methods.

Object detection

  • R-CNN - Applies convolutional features to bottom-up region proposals; transfers supervised pretraining to detection through domain-specific fine-tuning. Girshick et al. 2014.
    • Fast R-CNN - Builds on region-based convolutional detection to classify object proposals more efficiently; studies training and inference improvements over R-CNN and SPPnet. Girshick 2015.
    • Faster R-CNN - Learns region proposals with a network that shares image features with the detector, replacing a separate proposal-generation algorithm. Ren et al. 2015.
  • YOLO - Predicts bounding boxes and class probabilities directly from a full image in one network evaluation, framing detection as joint regression rather than proposal classification. Redmon et al. 2016.
  • SSD - Predicts class scores and offsets for default boxes at multiple feature-map resolutions, handling different object scales without a separate proposal stage. Liu et al. 2016.
  • Feature pyramid networks - Combines a top-down pathway with lateral connections to construct semantically rich features at multiple resolutions for detecting objects at different scales. Lin et al. 2017.
  • RetinaNet and focal loss - Downweights well-classified examples in cross-entropy training so abundant easy background examples do not dominate a dense detector’s loss. Lin et al. 2017.
  • DETR - Treats detection as set prediction: learned object queries and bipartite matching produce unique predictions without anchor generation or non-maximum suppression. Carion et al. 2020.
    • Deformable DETR - Restricts attention to a small set of sampling points around each reference point, addressing DETR’s training convergence and high-resolution feature limitations. Zhu et al. 2021.
  • Grounding DINO - Combines a DINO object detector with language-guided queries and cross-modal fusion to detect objects specified by category names or referring expressions. Liu et al. 2024.

Semantic, instance, and prompted segmentation

  • Fully convolutional networks - Adapts classification networks to pixelwise prediction; combines deep coarse semantics with shallower spatial detail to produce segmentation maps for variable-sized images. Long et al. 2015.
  • U-Net - Combines a contracting context path with a symmetric expanding localization path. The biomedical study uses strong augmentation to train from limited annotated images. Ronneberger et al. 2015.
  • Mask R-CNN - Adds a parallel mask-prediction branch to Faster R-CNN, producing a separate segmentation mask for each detected object instance. He et al. 2017.
  • DeepLab - Uses atrous convolution and spatial pyramid pooling for multiscale semantic segmentation; a fully connected conditional random field refines object boundaries. Chen et al. 2018.
  • Mask2Former - Constrains cross-attention to predicted mask regions. A shared architecture supports semantic, instance, and panoptic segmentation through their different grouping targets. Cheng et al. 2022.
  • Segment Anything - Trains a promptable segmentation model alongside a model-assisted mask collection process; evaluates transfer to new distributions and tasks without task-specific training. Kirillov et al. 2023.
    • SAM 2 - Extends promptable segmentation to images and videos with a Transformer using streaming memory; couples model development with interactive video-mask collection. Ravi et al. 2024.

Video and motion

  • Two-stream networks - Combines spatial appearance and temporal motion networks for action recognition; the temporal stream learns from stacked dense optical-flow fields. Simonyan and Zisserman 2014.
  • I3D - Inflates 2D convolution and pooling kernels into 3D to learn video features while reusing image-model designs and parameters; introduces the Kinetics action dataset. Carreira and Zisserman 2017.
  • SlowFast - Pairs a low-frame-rate pathway for spatial semantics with a lightweight high-frame-rate pathway for motion, separating temporal resolution from channel capacity. Feichtenhofer et al. 2019.
  • RAFT - Builds multiscale all-pairs pixel correlation volumes and recurrently updates an optical-flow field by querying those correlations. Teed and Deng 2020.
  • CoTracker - Jointly tracks multiple 2D points with a transformer that models dependencies between tracks, using windowed inference to extend trajectories through video and occlusions. Karaev et al. 2024.
  • AdaCodec - Encodes video as keyframes plus motion/residual tokens. At the same 2 FPS, retains benchmark accuracy with about one-seventh of RGB visual tokens on Qwen3-VL-8B. Equal-token comparisons instead allocate the savings to denser video coverage. Hou et al. 2026.

Geometry and scene representations

  • Structure-from-Motion revisited - Develops an incremental reconstruction pipeline for unordered image collections, addressing robustness, accuracy, completeness, and scalability; releases the pipeline as open-source software. Schonberger and Frahm 2016.
  • PointNet - Processes point clouds directly while respecting invariance to input-point ordering; supports object classification, part segmentation, and scene labeling without voxelizing the input. Qi et al. 2017.
  • NeRF - Fits a continuous field of volume density and view-dependent radiance to posed images; differentiable volume rendering synthesizes new views of the fitted scene. Mildenhall et al. 2020.
  • Instant neural graphics primitives - Augments a small neural network with trainable multiresolution hash-table features; optimized GPU kernels support efficient fitting and evaluation of neural graphics representations. Müller et al. 2022.
  • 3D Gaussian splatting - Represents a scene with optimized anisotropic Gaussians, interleaves parameter fitting with density control, and renders through visibility-aware splatting. Kerbl et al. 2023.
    • 2D Gaussian splatting - Represents surfaces with oriented planar Gaussian disks; ray-splat intersections and depth/normal regularization address geometry inconsistency in volumetric Gaussian representations. Huang et al. 2024.
  • DUSt3R - Predicts paired pointmaps without supplied camera calibration or poses; aligns these maps into a common frame for reconstruction from multiple images. Wang et al. 2024.
  • Depth Anything V2 - Trains a depth teacher on synthetic labels, then students on pseudo-labeled real images. Produces relative monocular depth; metric-depth models require separate fine-tuning. Yang et al. 2024.
  • VGGT - Uses a feed-forward network to infer camera parameters, depth, pointmaps, and point tracks from scene views, combining several geometric prediction tasks in one model. Wang et al. 2025.
  • DéjàView - Loops a shared Transformer block for multi-view reconstruction. Controlled ablations improve over separate blocks at each step; time and state gates add gains. The learned recurrence fails beyond its trained step range. Burzio et al. 2026.

Evaluation and dataset shift

  • COCO - Introduces everyday scenes with object-instance segmentation annotations and detection baselines, moving recognition evaluation toward objects in context. Lin et al. 2014.
  • Perceptual similarity and LPIPS - Evaluates deep-feature distances against human judgments of image similarity. Results concern the studied perceptual comparisons; classification accuracy alone is not the evaluation target. Zhang et al. 2018.
  • LVIS - Evaluates large-vocabulary instance segmentation with a long tail of rare categories. Category frequency exposes failures hidden by aggregate accuracy; the original paper motivates evaluation where many categories have few labeled instances. Gupta et al. 2019.
  • ImageNet-C and ImageNet-P - Benchmark image classifiers under common corruptions and perturbations. These tests measure a different robustness setting from worst-case adversarial attacks. Hendrycks and Dietterich 2019.
  • ImageNet replication - Recreates CIFAR-10 and ImageNet test sets and observes accuracy drops despite similar collection procedures; probes generalization beyond repeatedly reused benchmark images. Recht et al. 2019.
  • TIDE detection diagnostics - Separates six detection/segmentation error types and estimates their effects on average precision. Classification, localization and missed detections can yield the same headline score; the decomposition uses counterfactual error corrections. Bolya et al. 2020.
  • Dynamic view synthesis evaluation - Shows how monocular-video protocols can contain effective multiview signals; proposes co-visibility-aware image metrics and correspondence evaluation to expose this dependence. Gao et al. 2022.
  • ImageNet reannotation - Revises class definitions and adds multiple object labels. GPT-5.4 gains 5.10 accuracy points, versus 0.05 for EVA-02. The comparison excludes images with no valid class; annotation choices affect model families differently. Volkov et al. 2026.
  • Detection annotation errors - Reannotation of KITTI, COCO, Cityscapes and VOC finds missing boxes dominate label errors. Scores shift while model rankings remain largely stable. Soft labels use at least 11 annotators per object; evaluated subsets and classes differ by dataset. Penquitt et al. 2026.
  • DD-Ranking - Separates image content from teacher soft labels and augmentation. On ImageNet-1k, random subsets with the same labeler beat most tested distilled sets. Composite scores measure selected robustness criteria, not universal dataset quality. Li et al. 2026.