object_detection

Contents

• 1) Pubilc Datasets and Challenges
• 2) Annotation Tools
• 3) Pioneers and Experts
• 4) Blogs and Videos
  • ▶ GitHub
  • ▶ CSDN or zhihu
• 5) Papers and Sources Codes
  • ▶ Related Survey
  • ▶ Two-stage Anchor based
  • ▶ One-stage Anchor based
  • ▶ One-stage Anchor free
  • ▶ Detection Transformer (DETR)
  • ▶ YOLO Series Algorithms

1) Pubilc Datasets and Challenges

• (CVPR 2016) Cityscapes Dataset [github - CityscapesScripts]
• (CVPR 2016) TT100K (Traffic-Sign Detection and Classification in the Wild)(中国交通信号标志数据集)
• (arxiv 2018) CrowdHuman Dataset (A Benchmark for Detecting Human in a Crowd)(拥挤人群人体检测)
• (TPAMI 2023) SODA: A large-scale Small Object Detection dAtaset (小目标检测) [method CFINet]

2) Annotation Tools

• CSAILVision/LabelMeAnnotationTool [Source code for the LabelMe annotation tool.]
• tzutalin/LabelImg [🖍️ LabelImg is a graphical image annotation tool and label object bounding boxes in images]
• wkentaro/Labelme [Image Polygonal Annotation with Python (polygon, rectangle, circle, line, point and image-level flag annotation)]
• openvinotoolkit/CVAT [A Powerful and efficient Computer Vision Annotation Tool]
• Ericsson/EVA [A web-based tool for efficient annotation of videos and image sequences and has an additional tracking capabilities]

3) Pioneers and Experts

👍Martin Danelljan

4) Blogs and Videos

▶ GitHub

• (github) High-resolution networks (HRNets) for object detection
• (github) MMDetection: an open source object detection toolbox based on PyTorch by CUHK
• (github) TensorFlow Object Detection API
• (github) Yolov5 Yolov4 Yolov3 TensorRT Implementation
• (github) A Faster Pytorch Implementation of Faster R-CNN
• (github) FAIR's research platform for object detection research, implementing popular algorithms
• (github) Awesome Detection Transformer for Computer Vision (CV) (by IDEA-Research)
• (github) detrex is a research platform for Transformer-based Instance Recognition algorithms (by IDEA-Research)
• (github) (yolo-gradcam) yolo model with gradcam visual. 即插即用,不需要对源码进行任何修改!
• 👍(github) A collection of some awesome public YOLO object detection series projects

▶ CSDN or zhihu

• (CSDN blog) 理解COCO的评价指标：AP，AP50，AP70，mAP，AP[.50:.05:.95]
• (CSDN blog) 目标检测——One-stage和Two-stage的详解
• (CSDN blog) Anchor-free的目标检测文章
• (CSDN blog) 目标检测Anchor-free分支：基于关键点的目标检测（最新网络全面超越YOLOv3）
• (CSDN blog) YOLO V4 Tiny改进版来啦！速度294FPS精度不减YOLO V4 Tiny
• (blog) Custom Object Detection Tutorial with YOLO V5
• (zhihu) 如何评价YOLOv5？

5) Papers and Sources Codes

▶ Related Survey

• (ACM Computing Surveys 2022.09) Transformers in Vision: A Survey [paper link][arxiv link (2021.01)][Australian, USA]

• (ACM Computing Surveys 2022.12) Efficient Transformers: A Survey [paper link][arxiv link (2020.09)][Google]

• (AI Open 2022) A Survey of Transformers [paper link][arxiv link (2021.06][FDU]

• (TPAMI 2022) A Survey on Vision Transformer [paper link][arxiv link (2020.12)][Huawei, PKU, Dacheng Tao]

• (TPAMI 2023) Towards Large-Scale Small Object Detection: Survey and Benchmarks [paper link][arxiv link (2022.07)][project link][datasets SODA-D and SODA-A]

▶ Two-stage Anchor based

• ❤R-FCN(NIPS2016) R-FCN: Object Detection via Region-based Fully Convolutional Networks [arxiv link][Codes|Caffe&MATLAB(offical)][Codes|Caffe(unoffical)]

• DCN(ICCV2017) Deformable Convolutional Networks [arxiv link][Codes|MXNet(offical based on R-FCN)][Codes|MXNet(unoffical based on R-FCN)]

• ❤MaskRCNN(ICCV2017) Mask R-CNN [paper link][codes|official]

▶ One-stage Anchor based

• RetinaNet(ICCV2017) Focal Loss for Dense Object Detection [arxiv link][Codes|PyTorch(unoffical)][[Codes|Keras(unoffical)]

• Repulsion(CVPR2018) Repulsion Loss: Detecting Pedestrians in a Crowd [arxiv link][Codes|PyTorch(unoffical using SSD)][Codes|PyTorch(unoffical using RetinaNet)][CSDN blog]

▶ One-stage Anchor free

• CornerNet(ECCV2018) CornerNet: Detecting Objects as Paired Keypoints [arxiv link][Codes|PyTorch(offical)][Codes|PyTorch(offical CornerNet-Lite)]

• ❤CenterNet(arxiv2019) Objects as Points [arxiv link][Codes|PyTorch(offical)]

• CenterNet(arxiv2019) CenterNet: Keypoint Triplets for Object Detection [arxiv link][Codes|PyTorch(offical)]

• ❤FCOS(ICCV2019) FCOS: Fully Convolutional One-Stage Object Detection [arxiv link][Codes|PyTorch_MASK_RCNN(offical)][Codes|PyTorch(unoffical improved)][Codes|PyTorch(unoffical using HRNet as backbone)][blog_zhihu]

• VFNet(arxiv2020) VarifocalNet: An IoU-aware Dense Object Detector [arxiv link][Codes|offical with MMDetection & PyTorch]

• 👍ATSS(CVPR2020 Oral, Best Paper Nomination) Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection [paper link][arxiv link][code|official][CSDN blog][Shifeng Zhang]

• GFLV1 (NIPS2020) Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection [paper link][arxiv link][code|official][Li Xiang, based on ATSS ]

• GFLV2 (CVPR2021) Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection [paper link][arxiv link][code|official][Li Xiang, based on ATSS ]

• 👍TOOD(ICCV2021 Oral) TOOD: Task-aligned One-stage Object Detection [paper link][arxiv link][project link][code|official]

• GFL(QFL + DFL)(TPAMI2022) Generalized Focal Loss: Towards Efficient Representation Learning for Dense Object Detection [paper link][Li Xiang, based on ATSS ]

▶ Detection Transformer (DETR, fully end-to-end)

DETRs --> Pros: eliminates the hand-designed anchor and NMS components; Cons: slow training convergence and hard-to-optimize queries. Thus, researchers should put their efforts of optimizing transformer-based detectors in accelerating training convergence and reducing optimization difficulty

• ❤DEtection TRansformer(DETR) (ECCV2020 BestPaper) End-to-End Object Detection with Transformers [paper link][codes|official][bilibili paper reading video][bilibili paper reading video 2][facebookresearch]

• ❤DeformableDETR (ICLR2021 OralPaper) Deformable DETR: Deformable Transformers for End-to-End Object Detection [paper link] [codes|official][bilibili paper reading video]

• ConditionalDETR (ICCV2021) Conditional DETR for Fast Training Convergence [paper link][arxiv link][code|official]

• DeFCN (CVPR2021) End-to-End Object Detection with Fully Convolutional Network [paper link][arxiv link][codes|official][Megvii BaseDetection]

• SparseR-CNN (CVPR2021, TPAMI2023) Sparse R-CNN: End-to-End Object Detection with Learnable Proposals [paper link][arxiv link][code|official]

• Anchor-DETR (AAAI2022) Anchor detr: Query design for transformer-based detector [paper link][Megvii Technology]

• SparseDETR (ICLR2022) Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity [openreview link][arxiv link][code|official]

• DAB-DETR (ICLR2022) DAB-DETR: Dynamic Anchor Boxes are Better Queries for DETR [paper link][codes|official]

• DN-DETR (CVPR2022 OralPaper) DN-DETR: Accelerate DETR Training by Introducing Query DeNoising [paper link] [codes|official][IDEA Research, https://www.idea.edu.cn/]

• DINO (ICLR2023) DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection [openreview link][arxiv link][codes|official][IDEA Research]

• HDETR (CVPR2023) DETRs with Hybrid Matching [arxiv link][code|official][code|official (H-Deformable-DETR)][code|official (H-PETR-3D)][code|official (H-PETR-Pose)]

• DDQ-DETR (CVPR2023) Dense Distinct Query for End-to-End Object Detection [arxiv link][code|official]

• LiteDETR (CVPR2023) Lite DETR : An Interleaved Multi-Scale Encoder for Efficient DETR [paper link][arxiv link][code|official][IDEA Research, https://www.idea.edu.cn/]

• ❤RT-DETR(CVPR2024)(arxiv2023) DETRs Beat YOLOs on Real-time Object Detection [paper link][arxiv link][code|official whole-repo][code|official branch-repo][code|reproduced by ultralytics][Baidu, PaddlePaddle, PaddleDetection]

▶ YOLO Series Algorithms

Anchor-base: YOLOv2, YOLOv3, YOLOv4, Scaled-YOLOv4, YOLOv5, YOLOR, TPH-YOLOv5, YOLOv5-Lite, YOLOv7; Anchor-Free: YOLOv1, YOLOX, PP-YOLOE, YOLOv5u, YOLOv6, YOLOv7u, YOLOv8

• ❤YOLOv1(CVPR2016) You Only Look Once: Unified, Real-Time Object Detection [paper link][project link (darknet)][code|official by darknet][Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi]

• ❤YOLOv2(CVPR2017) YOLO9000: Better, Faster, Stronger [paper link][project link (darknet)][code|unofficial PyTorch][Joseph Redmon, Ali Farhadi]

• YOLOv3(arxiv2018.04) YOLOv3: An Incremental Improvement [paper link][project link (darknet)][code|unofficial PyTorch (github 2018.12)]code|unofficial PyTorch][paper: Joseph Redmon, Ali Farhadi][code: ultralytics, Glenn Jocher]

• YOLOv4(arxiv2020.04) YOLOv4: Optimal Speed and Accuracy of Object Detection [paper link][code|official DarkNet][code|official PyTorch_YOLOv4][code|unofficial PyTorch][Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao, Taiwan + Intel, DarkNet]

• Scaled-YOLOv4(CVPR2021) Scaled-YOLOv4: Scaling Cross Stage Partial Network [paper link][code|official][Chien-Yao Wang, Alexey Bochkovskiy, Hong-Yuan Mark Liao, Taiwan + Intel]

• ❤YOLOv5(github 2020.06) YOLOv5 🚀 in PyTorch > ONNX > CoreML > TFLite [No published paper][code|official by ultralytics][Docs|official by ultralytics][author: Glenn Jocher][ultralytics, Glenn Jocher]

• YOLOR(arxiv2021.05) You Only Learn One Representation: Unified Network for Multiple Tasks [paper link][code|official][Chien-Yao Wang, I-Hau Yeh, Hong-Yuan Mark Liao, Taiwan]

• ❤YOLOX(arxiv2021.07) YOLOX: Exceeding YOLO Series in 2021 [paper link][code|official by MegEngine-BaseDetection][code|official by MegEngine][Docs|official][Megvii, Jian Sun]

• TPH-YOLOv5(ICCVW2021) TPH-YOLOv5: Improved YOLOv5 Based on Transformer Prediction Head for Object Detection on Drone-Captured Scenarios [paper link][arxiv link][code|official PyTorch][Transformer, SwinTransformer]

• YOLOv5-Lite(github 2021.08) YOLOv5-Lite：Lighter, faster and easier to deploy [No published paper][code|official]

• PP-YOLOE(arxiv2022.03) PP-YOLOE: An evolved version of YOLO [paper link][code|official][PaddlePaddle, PaddleDetection, Baidu, based on the TAL (task-aligned assignment loss) proposed in TOOD]

• YOLOv6(arxiv2022.09) YOLOv6: A Single-Stage Object Detection Framework for Industrial Applications [paper link][code|official by meituan][YOLOv6 v3.0: A Full-Scale Reloading (arxiv2023.01)][meituan]

• ❤YOLOv7(CVPR2023)(arxiv2022.07) YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors [paper link][code|official][code|official - pose , blog - ipynb][code|official - seg, blog - ipynb][code|YOLOv7-Seg based on YOLOv5][code|YOLOv7-Pose-Estimation by RizwanMunawar][code|YOLOv7-Segmentation by RizwanMunawar][blog|YOLOv7-Segmentation by RizwanMunawar][Chien-Yao Wang, Alexey Bochkovskiy, Hong-Yuan Mark Liao, Taiwan + Intel][re-parameterized module + dynamic label assignment + trainable bag-of-freebies][Following the YOLOv5 project arch]

• DAMO-YOLO(arxiv2022.11) DAMO-YOLO: A Report on Real-Time Object Detection Design [arxiv link][project link (github)]

• RTMDet(arxiv2022.12) RTMDet: An Empirical Study of Designing Real-Time Object Detectors [arxiv link][code|official][Shanghai AI Laboratory + S-Lab, Nanyang Technological University + Tianjin University + SJTU]

• ❤YOLOv8(github 2023.01) YOLOv8 🚀 in PyTorch > ONNX > CoreML > TFLite [No published paper][code|official by ultralytics][Docs|official by ultralytics][ultralytics, Glenn Jocher]

• YOLO-NAS(github 2023.05) A Next-Generation, Object Detection Foundational Model generated by Deci’s Neural Architecture Search Technology [project link][super-gradients github][zhihu blog]

• YOLO-MS(arxiv2023.08) YOLO-MS: Rethinking Multi-Scale Representation Learning for Real-time Object Detection [arxiv link][code|official][Nankai University]

• Gold-YOLO(NIPS2023)(arxiv2023.09) Gold-YOLO: Efficient Object Detector via Gather-and-Distribute Mechanism [paper link][arxiv link][code|official][Huawei]

• YOLOv9(arxiv2024.02) YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information [arxiv link][code|official][Chien-Yao Wang, I-Hau Yeh and Hong-Yuan Mark Liao in Taiwan]

• YOLOv10(arxiv2024.05) YOLOv10: Real-Time End-to-End Object Detection [arxiv link][code|official][THU][The code base is built with ultralytics and RT-DETR.]