YOLO9000: Better, Faster, Stronger
By Joseph Redmon, Ali Farhadi. 2016
17,782 citations recorded by the source corpus.
We introduce YOLO9000, a state-of-the-art, real-time object detection system that can detect over 9000 object categories. First we propose various improvements to the YOLO detection method, both novel and drawn from prior work. The improved model, YOLOv2, is state-of-the-art on standard detection tasks like PASCAL VOC and COCO. Using a novel, multi-scale training method the same YOLOv2 model can run at varying sizes, offering an easy tradeoff between speed and accuracy. At 67 FPS, YOLOv2 gets 76.8 mAP on VOC 2007. At 40 FPS, YOLOv2 gets 78.6 mAP, outperforming state-of-the-art methods like Faster RCNN with ResNet and SSD while still running significantly faster. Finally we propose a method to jointly train on object detection and classification. Using this method we train YOLO9000 simultaneously on the COCO detection dataset and the ImageNet classification dataset. Our joint training allows YOLO9000 to predict detections for object classes that dont have labelled detection data. We validate our approach on the ImageNet detection task. YOLO9000 gets 19.7 mAP on the ImageNet detection validation set despite only having detection data for 44 of the 200 classes. On the 156 classes not in COCO, YOLO9000 gets 16.0 mAP. YOLO9000 predicts detections for more than 9000 different object categories, all in real-time.
Subfields: Computer Vision and Pattern Recognition
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Computer Vision and Pattern Recognition
- Luc Van Gool
- Dacheng Tao
- Xiaogang Wang
- Ming-Hsuan Yang
- Trevor Darrell
- Jitendra Malik
- Andrew Zisserman
- Bernt Schiele
Citations in the corpus
- A Multimodal Framework for the Detection of Hateful Memes (2020, background): “We also find that margin ranking loss (UNITER CV15 + CFU + HW + MRL ) and YOLO9000 [28] object tags (UNITER CV15 + CFU + HW + YOLO ) do not improve Phase 1 performance.”
- Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive Training Sample Selection (2019, neutral): “anchor-based one-stage: YOLOv2 [46] trainval35k DarkNet-19 21.”
- Objects as Points (2019, background): “Several improvements to one-stage detectors include anchor shape priors [44, 45], different feature resolution [36], and loss re-weighting among different samples [33].”
- FCOS: Fully Convolutional One-Stage Object Detection (2019, methodology): “One-stage methods: YOLOv2 [22] DarkNet-19 [22] 21.”
- Generalized Intersection Over Union: A Metric and a Loss for Bounding Box Regression (2019, background): “Most popular object detectors [20, 21, 3, 12, 13, 16] utilize some combination of the bounding box representations and losses mentioned above.”
- Deep Perm-Set Net: Learn to predict sets with unknown permutation and cardinality using deep neural networks (2018, methodology): “We compare it with the state-of-the art object detectors, i.e. Faster-RCNN Ren et al. (2015) and YOLO v2 Redmon & Farhadi (2017).”
- Speed/Accuracy Trade-Offs for Modern Convolutional Object Detectors (2016, methodology): “[32, 28, 29, 9] use SSD-like architectures to predict final (1 of K) class labels.”
- ORBIT: A Real-World Few-Shot Dataset for Teachable Object Recognition (2021, background): “Object recognition systems have made spectacular advances in recent years [42, 47, 43, 37, 14, 30, 36] however, most systems still rely on training datasets with 100s to 1,000s of high-quality, labeled examples per object category.”
- Towards Real-World Prohibited Item Detection: A Large-Scale X-ray Benchmark (2021, methodology): “After that, different one-stage object detectors are proposed, including SSD [24], DSSD [8], and YOLO series [28, 29, 30].”
- Learning Open-World Object Proposals without Learning to Classify (2021, background): “Approaches under semisupervised setup [29], [30], [31], [32] employ the weak imagelevel labels for novel classes as well as box-level labels for base classes.”
- MobileNetV2: Inverted Residuals and Linear Bottlenecks (2018, neutral): “We also compare to YOLOv2 [35] and original SSD (with VGG-16 [6] as base network) as baselines.”
- Bilinear Attention Networks (2018, neutral): “…RCNN [7] 43.89 64.46 68.66 76.91 Rohrbach et al. [26] Fast RCNN [7] 48.38 - - 77.90 Fukui et al. [6] Fast RCNN [7] 48.69 - - - Plummer et al. [23] Fast RCNN [7] † 50.89 71.09 75.73 85.12 Yeh et al. [37] YOLOv2 [24] ‡ 53.97 - - - Hinami and Satoh [10] Query-Adaptive RCNN [10] multi-channel input.”
- The Open Images Dataset V4 (2018, methodology): “More recently, single-shot detectors were presented to bypass the computational bottleneck of object proposals by regressing object locations directly from a predefined set of anchor boxes (e.g. SSD (Liu et al., 2016) and YOLO (Red-mon et al., 2016; Redmon and Farhadi, 2017)).”
- VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection (2017, background): “Region proposal network (RPN) [34] is a highly optimized algorithm for efficient object detection [17, 5, 33, 24].”