Image Classification on Birdsnap (test)
85.5Top-1 AccJFT - Bird
Evaluation Results
| Method | Links | |
|---|---|---|
| JFT - BirdBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=JFT (Hand-selected subset)2018.11 | 85.5 | |
| Gradient-based FVEpart detector=CS-Parts [24], optimization=gradient descent2020.07 | 85.3 | |
| JFT - Adaptive TransferBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=JFT2018.11 | 85.1 | |
| EM-based FVEpart detector=CS-Parts [24], optimization=online EM algorithm2020.07 | 84.9 | |
| FixSENet-1542020.07 | 84.3 | |
| JFT - AnimalBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=JFT (Hand-selected subset)2018.11 | 84.1 | |
| GAPpart detector=CS-Parts [24], aggregation=concatenated part features and GAP2020.07 | 84 | |
| Best Published Result [17]Protocol=Reference value2018.11 | 83.9 | |
| GPipe (AmoebaNet-B (18, 512))Architecture=AmoebaNet-B (18, 512), Resolution=480x480, Protocol=Fine-tuned, Number of fine-tuning runs=5, Crop=Single-crop2018.11 | 83.6 | |
| No Parts (baseline)parts=none, features=global image only2020.07 | 81.9 | |
| ImageNet - Entire DatasetBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=ImageNet2018.11 | 80.8 | |
| ImageNet - Adaptive TransferBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=ImageNet2018.11 | 80.7 | |
| Entire JFT DatasetBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=JFT2018.11 | 80.3 | |
| Method [32]2018.11 | 80.2 | |
| JFT - FoodBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=JFT (Hand-selected subset)2018.11 | 79.7 | |
| JFT - TransportBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=JFT (Hand-selected subset)2018.11 | 79.2 | |
| JFT - CarBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=JFT (Hand-selected subset)2018.11 | 79 | |
| JFT - VehicleBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=JFT (Hand-selected subset)2018.11 | 78.8 | |
| JFT - AircraftBackbone=AmoebaNet-B, Input Resolution=331 x 331, Pre-training Source=JFT (Hand-selected subset)2018.11 | 78 | |
| ReLICv2Evaluation protocol=Fine-tuned, Backbone=ResNet-50, Pre-trained=ImageNet2022.01 | 76.7 | |
| BYOLEvaluation Protocol=Fine-tuned2021.06 | 76.3 | |
| BYOLEvaluation protocol=Fine-tuned, Backbone=ResNet-50, Pre-trained=ImageNet2022.01 | 76.3 | |
| Random initEvaluation Protocol=Fine-tuned2021.06 | 76.1 | |
| Random InitEvaluation protocol=Fine-tuned, Backbone=ResNet-50, Pre-trained=None2022.01 | 76.1 | |
| SimCLREvaluation Protocol=Fine-tuned2021.06 | 75.9 | |
| SimCLREvaluation protocol=Fine-tuned, Backbone=ResNet-50, Pre-trained=ImageNet2022.01 | 75.9 | |
| Supervised-INEvaluation Protocol=Fine-tuned2021.06 | 75.8 | |
| Supervised-INEvaluation protocol=Fine-tuned, Backbone=ResNet-50, Pre-trained=ImageNet2022.01 | 75.8 | |
| SSL-HSIC (w/ target)Evaluation Protocol=Fine-tuned2021.06 | 74.9 | |
| SSL-HSIC (w/o target)Evaluation Protocol=Fine-tuned2021.06 | 73.1 | |
| ReLICv2Evaluation protocol=Linear evaluation, Backbone=ResNet-50, Pre-trained=ImageNet2022.01 | 65.4 | |
| NNCLREvaluation protocol=Linear evaluation, Backbone=ResNet-50, Pre-trained=ImageNet2022.01 | 61.4 | |
| SSL-HSIC (w/ target)Evaluation Protocol=Linear2021.06 | 57.8 | |
| BYOLEvaluation Protocol=Linear2021.06 | 57.2 | |
| BYOLEvaluation protocol=Linear evaluation, Backbone=ResNet-50, Pre-trained=ImageNet2022.01 | 57.2 | |
| Supervised-INEvaluation Protocol=Linear2021.06 | 53.7 | |
| Supervised-INEvaluation protocol=Linear evaluation, Backbone=ResNet-50, Pre-trained=ImageNet2022.01 | 53.7 | |
| SSL-HSIC (w/o target)Evaluation Protocol=Linear2021.06 | 50.6 | |
| ReLICv2Backbone=ResNet-50, Pre-training dataset=JFT, Pre-training epochs=5000 ImageNet-equivalent, Evaluation protocol=Linear transfer2022.01 | 49.4 | |
| ReLICv2Backbone=ResNet-50, Pre-training dataset=JFT, Pre-training epochs=1000 ImageNet-equivalent, Evaluation protocol=Linear transfer2022.01 | 47.4 | |
| DnCBackbone=ResNet-50, Pre-training dataset=JFT, Pre-training epochs=4500 ImageNet-equivalent, Evaluation protocol=Linear transfer2022.01 | 42.1 | |
| BYOLBackbone=ResNet-50, Pre-training dataset=JFT, Pre-training epochs=5000 ImageNet-equivalent, Evaluation protocol=Linear transfer2022.01 | 38.2 | |
| SimCLREvaluation Protocol=Linear2021.06 | 37.4 | |
| SimCLREvaluation protocol=Linear evaluation, Backbone=ResNet-50, Pre-trained=ImageNet2022.01 | 37.4 |