Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
Deeper neural networks are more difficult to train. We present a residual learning framework to ease the training of networks that are substantially deeper than those used previously. We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, instead of learning unreferenced functions. We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth. On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers---8x deeper than VGG nets but still having lower complexity. An ensemble of these residual nets achieves 3.57% error on the ImageNet test set. This result won the 1st place on the ILSVRC 2015 classification task. We also present analysis on CIFAR-10 with 100 and 1000 layers. The depth of representations is of central importance for many visual recognition tasks. Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset. Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Autonomous Vehicles | UAV-Human | Backpack | 63.5 | ResNet |
| Autonomous Vehicles | UAV-Human | Gender | 74.7 | ResNet |
| Autonomous Vehicles | UAV-Human | Hat | 65.2 | ResNet |
| Autonomous Vehicles | UAV-Human | LCC | 49.7 | ResNet |
| Autonomous Vehicles | UAV-Human | LCS | 69.3 | ResNet |
| Autonomous Vehicles | UAV-Human | UCC | 44.4 | ResNet |
| Autonomous Vehicles | UAV-Human | UCS | 68.9 | ResNet |
| Image-to-Image Translation | GTAV-to-Cityscapes Labels | mIoU | 41.7 | ResNet101 65.1 |
| Image-to-Image Translation | Syn2Real-C | Accuracy | 52.4 | No Adaptation |
| Domain Adaptation | Office-31 | Average Accuracy | 76.1 | ResNet-50 |
| Domain Adaptation | Office-Home | Accuracy | 59.9 | ResNet-50 [cite:CVPR16DRL] |
| Domain Adaptation | ImageNet-R | Top-1 Error Rate | 63.9 | ResNet-50 |
| Domain Adaptation | ImageNet-A | Top-1 accuracy % | 4.2 | ResNet-50 (300 Epochs) |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 47.5 | ResNet-152 |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 51.3 | ResNet-152 |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 43.3 | ResNet-152 |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 46.3 | ResNet-101 |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 50.1 | ResNet-101 |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 40.5 | ResNet-101 |
| Domain Adaptation | VizWiz-Classification | Accuracy - All Images | 42.9 | ResNet-50 |
| Domain Adaptation | VizWiz-Classification | Accuracy - Clean Images | 47.7 | ResNet-50 |
| Domain Adaptation | VizWiz-Classification | Accuracy - Corrupted Images | 37.1 | ResNet-50 |
| Image Generation | GTAV-to-Cityscapes Labels | mIoU | 41.7 | ResNet101 65.1 |
| Image Generation | Syn2Real-C | Accuracy | 52.4 | No Adaptation |
| Person Re-Identification | SYSU-30k | Rank-1 | 20.1 | ResNet-50 (generalization) |
| Crowds | UCF-QNRF | MAE | 190 | Resnet101 |
| Speaker Verification | VoxCeleb2 | EER | 100 | ResNet-50 |
| Semantic Segmentation | Cityscapes val | mIoU | 75.7 | Dilated-ResNet (Dilated-ResNet-101) |
| Semantic Segmentation | DADA-seg | mIoU | 23.6 | ResNet-101 |
| Semantic Segmentation | DADA-seg | mIoU | 18.96 | ResNet-50 |
| Multi-Label Image Classification | VizWiz-Classification | Accuracy | 47.5 | ResNet151 |
| Pedestrian Attribute Recognition | UAV-Human | Backpack | 63.5 | ResNet |
| Pedestrian Attribute Recognition | UAV-Human | Gender | 74.7 | ResNet |
| Pedestrian Attribute Recognition | UAV-Human | Hat | 65.2 | ResNet |
| Pedestrian Attribute Recognition | UAV-Human | LCC | 49.7 | ResNet |
| Pedestrian Attribute Recognition | UAV-Human | LCS | 69.3 | ResNet |
| Pedestrian Attribute Recognition | UAV-Human | UCC | 44.4 | ResNet |
| Pedestrian Attribute Recognition | UAV-Human | UCS | 68.9 | ResNet |
| Object Detection | COCO minival | AP50 | 64.3 | Cascade Mask R-CNN (ResNet-50) |
| Object Detection | COCO minival | AP75 | 50.5 | Cascade Mask R-CNN (ResNet-50) |
| Object Detection | COCO minival | box AP | 46.3 | Cascade Mask R-CNN (ResNet-50) |
| Object Detection | COCO minival | AP50 | 63 | GFL (ResNet-50) |
| Object Detection | COCO minival | AP75 | 48.3 | GFL (ResNet-50) |
| Object Detection | COCO minival | box AP | 44.5 | GFL (ResNet-50) |
| Object Detection | COCO minival | AP50 | 61.9 | ATSS (ResNet-50) |
| Object Detection | COCO minival | AP75 | 47 | ATSS (ResNet-50) |
| Object Detection | COCO minival | box AP | 43.5 | ATSS (ResNet-50) |
| Image Classification | GasHisSDB | Accuracy | 98.56 | ResNet-50 |
| Image Classification | GasHisSDB | F1-Score | 99.24 | ResNet-50 |
| Image Classification | GasHisSDB | Precision | 99.94 | ResNet-50 |
| Image Classification | GasHisSDB | Accuracy | 98.47 | ResNet-18 |
| Image Classification | GasHisSDB | F1-Score | 99.19 | ResNet-18 |
| Image Classification | GasHisSDB | Precision | 99.94 | ResNet-18 |
| Image Classification | OmniBenchmark | Average Top-1 Accuracy | 37.4 | ResNet-101 |
| Image Classification | OmniBenchmark | Average Top-1 Accuracy | 34.3 | ResNet-50 |
| Image Classification | cifar100 | 1:1 Accuracy | 45.98 | shreynet |
| Image Classification | ImageNet | GFLOPs | 11.3 | ResNet-152 |
| Image Classification | ImageNet | GFLOPs | 7.6 | ResNet-101 |
| Image Classification | ImageNet | GFLOPs | 3.8 | ResNet-50 |
| Image Classification | VizWiz-Classification | Accuracy | 47.5 | ResNet151 |
| 3D | COCO minival | AP50 | 64.3 | Cascade Mask R-CNN (ResNet-50) |
| 3D | COCO minival | AP75 | 50.5 | Cascade Mask R-CNN (ResNet-50) |
| 3D | COCO minival | box AP | 46.3 | Cascade Mask R-CNN (ResNet-50) |
| 3D | COCO minival | AP50 | 63 | GFL (ResNet-50) |
| 3D | COCO minival | AP75 | 48.3 | GFL (ResNet-50) |
| 3D | COCO minival | box AP | 44.5 | GFL (ResNet-50) |
| 3D | COCO minival | AP50 | 61.9 | ATSS (ResNet-50) |
| 3D | COCO minival | AP75 | 47 | ATSS (ResNet-50) |
| 3D | COCO minival | box AP | 43.5 | ATSS (ResNet-50) |
| Breast Tumour Classification | PCam | AUC | 0.948 | ResNet-50 (e) |
| Breast Tumour Classification | PCam | AUC | 0.942 | ResNet-34 (e) |
| Unsupervised Domain Adaptation | Office-Home | Accuracy | 59.9 | ResNet-50 [cite:CVPR16DRL] |
| 2D Classification | COCO minival | AP50 | 64.3 | Cascade Mask R-CNN (ResNet-50) |
| 2D Classification | COCO minival | AP75 | 50.5 | Cascade Mask R-CNN (ResNet-50) |
| 2D Classification | COCO minival | box AP | 46.3 | Cascade Mask R-CNN (ResNet-50) |
| 2D Classification | COCO minival | AP50 | 63 | GFL (ResNet-50) |
| 2D Classification | COCO minival | AP75 | 48.3 | GFL (ResNet-50) |
| 2D Classification | COCO minival | box AP | 44.5 | GFL (ResNet-50) |
| 2D Classification | COCO minival | AP50 | 61.9 | ATSS (ResNet-50) |
| 2D Classification | COCO minival | AP75 | 47 | ATSS (ResNet-50) |
| 2D Classification | COCO minival | box AP | 43.5 | ATSS (ResNet-50) |
| Classification | XImageNet-12 | Robustness Score | 0.8985 | ResNet 50 |
| Classification | NCT-CRC-HE-100K | Accuracy (%) | 94.72 | ResNet-50 |
| Classification | NCT-CRC-HE-100K | F1-Score | 97.09 | ResNet-50 |
| Classification | NCT-CRC-HE-100K | Precision | 100 | ResNet-50 |
| Classification | NCT-CRC-HE-100K | Specificity | 99.34 | ResNet-50 |
| Classification | NCT-CRC-HE-100K | Accuracy (%) | 92.66 | ResNet-18 |
| Classification | NCT-CRC-HE-100K | F1-Score | 95.23 | ResNet-18 |
| Classification | NCT-CRC-HE-100K | Precision | 99.9 | ResNet-18 |
| Classification | NCT-CRC-HE-100K | Specificity | 99.08 | ResNet-18 |
| 2D Object Detection | COCO minival | AP50 | 64.3 | Cascade Mask R-CNN (ResNet-50) |
| 2D Object Detection | COCO minival | AP75 | 50.5 | Cascade Mask R-CNN (ResNet-50) |
| 2D Object Detection | COCO minival | box AP | 46.3 | Cascade Mask R-CNN (ResNet-50) |
| 2D Object Detection | COCO minival | AP50 | 63 | GFL (ResNet-50) |
| 2D Object Detection | COCO minival | AP75 | 48.3 | GFL (ResNet-50) |
| 2D Object Detection | COCO minival | box AP | 44.5 | GFL (ResNet-50) |
| 2D Object Detection | COCO minival | AP50 | 61.9 | ATSS (ResNet-50) |
| 2D Object Detection | COCO minival | AP75 | 47 | ATSS (ResNet-50) |
| 2D Object Detection | COCO minival | box AP | 43.5 | ATSS (ResNet-50) |
| Medical Image Classification | NCT-CRC-HE-100K | Accuracy (%) | 94.72 | ResNet-50 |
| Medical Image Classification | NCT-CRC-HE-100K | F1-Score | 97.09 | ResNet-50 |
| Medical Image Classification | NCT-CRC-HE-100K | Precision | 100 | ResNet-50 |
| Medical Image Classification | NCT-CRC-HE-100K | Specificity | 99.34 | ResNet-50 |
| Medical Image Classification | NCT-CRC-HE-100K | Accuracy (%) | 92.66 | ResNet-18 |
| Medical Image Classification | NCT-CRC-HE-100K | F1-Score | 95.23 | ResNet-18 |
| Medical Image Classification | NCT-CRC-HE-100K | Precision | 99.9 | ResNet-18 |
| Medical Image Classification | NCT-CRC-HE-100K | Specificity | 99.08 | ResNet-18 |
| Domain Generalization | ImageNet-R | Top-1 Error Rate | 63.9 | ResNet-50 |
| Domain Generalization | ImageNet-A | Top-1 accuracy % | 4.2 | ResNet-50 (300 Epochs) |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 47.5 | ResNet-152 |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 51.3 | ResNet-152 |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 43.3 | ResNet-152 |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 46.3 | ResNet-101 |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 50.1 | ResNet-101 |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 40.5 | ResNet-101 |
| Domain Generalization | VizWiz-Classification | Accuracy - All Images | 42.9 | ResNet-50 |
| Domain Generalization | VizWiz-Classification | Accuracy - Clean Images | 47.7 | ResNet-50 |
| Domain Generalization | VizWiz-Classification | Accuracy - Corrupted Images | 37.1 | ResNet-50 |
| 10-shot image generation | Cityscapes val | mIoU | 75.7 | Dilated-ResNet (Dilated-ResNet-101) |
| 10-shot image generation | DADA-seg | mIoU | 23.6 | ResNet-101 |
| 10-shot image generation | DADA-seg | mIoU | 18.96 | ResNet-50 |
| 16k | COCO minival | AP50 | 64.3 | Cascade Mask R-CNN (ResNet-50) |
| 16k | COCO minival | AP75 | 50.5 | Cascade Mask R-CNN (ResNet-50) |
| 16k | COCO minival | box AP | 46.3 | Cascade Mask R-CNN (ResNet-50) |
| 16k | COCO minival | AP50 | 63 | GFL (ResNet-50) |
| 16k | COCO minival | AP75 | 48.3 | GFL (ResNet-50) |
| 16k | COCO minival | box AP | 44.5 | GFL (ResNet-50) |
| 16k | COCO minival | AP50 | 61.9 | ATSS (ResNet-50) |
| 16k | COCO minival | AP75 | 47 | ATSS (ResNet-50) |
| 16k | COCO minival | box AP | 43.5 | ATSS (ResNet-50) |
| 1 Image, 2*2 Stitching | GTAV-to-Cityscapes Labels | mIoU | 41.7 | ResNet101 65.1 |
| 1 Image, 2*2 Stitching | Syn2Real-C | Accuracy | 52.4 | No Adaptation |