Browse › Computer Vision › Action Recognition › Something-Something V2
Something-Something V2 Benchmark (Action Recognition)
Action Recognition is a computer vision task that involves recognizing human actions in videos or images. The goal is to classify and categorize the actions being performed in the video or image into a predefined set of action classes.
In the video domain, it is an open question whether training an action classification network on a sufficiently large dataset, will give a similar boost in performance when applied to a different temporal task or dataset. The challenges of building video datasets has meant that most popular benchmarks for action recognition are small, having on the order of 10k videos.
Please note some benchmarks may be located in the Action Classification or Video Classification tasks, e.g. Kinetics-400.
The archive carries no text for this table; the description above is the archive's text for the task Action Recognition. archive 2025-07-28
Over time archive 2025-07-28
The chart needs JavaScript; the table below carries every value.
Direction inferred from the metric name, not from the archive: Top-1 Accuracy (higher is better), Top-5 Accuracy (higher is better). Not inferred (points only, no best-so-far line): Parameters, GFLOPs. Points are placed at the row's paper date; 123 of 123 rows carry one.
Results archive 2025-07-28
Archive rows end at the archive snapshot, 2025-07-28: no result published after that date is in this table. Rank is the archive's row order at that snapshot; not re-ranked here. Metric values are the archive's strings. Column headers sort the table in your browser; each row keeps its archive rank.
| Paper | Code | Ran Syntology | Report | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | MVD (Kinetics400 pretrain, ViT-H, 16 frame) | 77.3 | 95.7 | 633 | 1192x6 | ✓ | Paper | Code | 2022 | linked, not harvested | report |
| 2 | DejaVid | 77.2 | 96.3 | ✓ | Paper | Code | 2025 | linked, not harvested | report | ||
| 3 | InternVideo | 77.2 | ✓ | Paper | Code | 2022 | 3 of 3 ran · 0 unverified | report | |||
| 4 | InternVideo2-1B | 77.1 | ✓ | Paper | Code | 2024 | linked, not harvested | report | |||
| 5 | VideoMAE V2-g | 77.0 | 95.9 | 1013 | 2544x6 | ✓ | Paper | Code | 2023 | 2 of 6 ran · 4 unverified | report |
| 6 | MVD (Kinetics400 pretrain, ViT-L, 16 frame) | 76.7 | 95.5 | 305 | 597x6 | ✓ | Paper | Code | 2022 | linked, not harvested | report |
| 7 | Hiera-L (no extra data) | 76.5 | – | Paper | Code | 2023 | 0 of 6 ran · 6 unverified | report | |||
| 8 | TubeViT-L | 76.1 | 95.2 | – | Paper | Code | 2022 | 1 of 3 ran · 2 unverified | report | ||
| 9 | VideoMAE (no extra data, ViT-L, 32x2) | 75.4 | 95.2 | 305 | 1436x3 | – | Paper | Code | 2022 | 9 of 13 ran · 4 unverified | report |
| 10 | Side4Video (EVA ViT-E/14) | 75.2 | 94.0 | – | Paper | Code | 2023 | linked, not harvested | report | ||
| 11 | MaskFeat (Kinetics600 pretrain, MViT-L) | 75.0 | 95.0 | 218 | 2828*3 | ✓ | Paper | Code | 2021 | linked, not harvested | report |
| 12 | MAR (50% mask, ViT-L, 16x4) | 74.7 | 94.9 | 311 | 276x6 | – | Paper | Code | 2022 | linked, not harvested | report |
| 13 | ATM | 74.6 | 94.4 | – | Paper | Code | 2023 | 3 of 6 ran · 3 unverified | report | ||
| 14 | MAWS (ViT-L) | 74.4 | ✓ | Paper | Code | 2023 | linked, not harvested | report | |||
| 15 | VideoMAE (no extra data, ViT-L, 16frame) | 74.3 | 94.6 | 305 | 597x6 | – | Paper | Code | 2022 | 9 of 13 ran · 4 unverified | report |
| 16 | MAR (75% mask, ViT-L, 16x4) | 73.8 | 94.4 | 311 | 131x6 | – | Paper | Code | 2022 | linked, not harvested | report |
| 17 | MVD (Kinetics400 pretrain, ViT-B, 16 frame) | 73.7 | 94.0 | 87 | 180x6 | ✓ | Paper | Code | 2022 | linked, not harvested | report |
| 18 | ViC-MAE (ViT-L) | 73.7 | – | Paper | Code | 2023 | linked, not harvested | report | |||
| 19 | TAdaFormer-L/14 | 73.6 | ✓ | Paper | Code | 2023 | 2 of 2 ran · 0 unverified | report | |||
| 20 | TDS-CLIP-ViT-L/14(8frames) | 73.4 | 93.8 | – | Paper | Code | 2024 | linked, not harvested | report | ||
| 21 | MViTv2-L (IN-21K + Kinetics400 pretrain) | 73.3 | 94.1 | 213.1 | – | Paper | Code | 2021 | linked, not harvested | report | |
| 22 | AMD(ViT-B/16) | 73.3 | 94.0 | 87 | 180x6 | – | Paper | – | 2023 | no code linked | report |
| 23 | UniFormerV2-L | 73.0 | 94.5 | 5154 | ✓ | Paper | Code | 2022 | linked, not harvested | report | |
| 24 | ST-Adapter (ViT-L, CLIP) | 72.3 | 93.9 | 8248 | ✓ | Paper | Code | 2022 | 4 of 10 ran · 6 unverified | report | |
| 25 | ZeroI2V ViT-L/14 | 72.2 | 93.0 | ✓ | Paper | Code | 2023 | 1 of 3 ran · 2 unverified | report | ||
| 26 | MViT-B (IN-21K + Kinetics400 pretrain) | 72.1 | 225x3 | ✓ | Paper | Code | 2021 | linked, not harvested | report | ||
| 27 | CAST(ViT-B/16) | 71.6 | – | Paper | Code | 2023 | 9 of 17 ran · 8 unverified | report | |||
| 28 | StructVit-B-4-1 | 71.5 | – | Paper | – | 2024 | no code linked | report | |||
| 29 | OMNIVORE (Swin-B, IN-21K+ Kinetics400 pretrain) | 71.4 | 93.5 | ✓ | Paper | Code | 2022 | 2 of 2 ran · 0 unverified | report | ||
| 30 | BEVT (IN-1K + Kinetics400 pretrain) | 71.4 | - | 89 | 321x3 | ✓ | Paper | Code | 2021 | linked, not harvested | report |
| 31 | UniFormer-B (IN-1K + Kinetics400 pretrain) | 71.2 | 92.8 | 50.1 | 259x3 | ✓ | Paper | Code | 2021 | linked, not harvested | report |
| 32 | TAdaConvNeXtV2-B | 71.1 | ✓ | Paper | Code | 2023 | 2 of 2 ran · 0 unverified | report | |||
| 33 | MAR (50% mask, ViT-B, 16x4) | 71.0 | 92.8 | 94 | 86x6 | – | Paper | Code | 2022 | linked, not harvested | report |
| 34 | MVD (Kinetics400 pretrain, ViT-S, 16 frame) | 70.9 | 92.8 | 22 | 57x6 | ✓ | Paper | Code | 2022 | linked, not harvested | report |
| 35 | CoVeR(JFT-3B) | 70.9 | 92.5 | ✓ | Paper | – | 2021 | no code linked | report | ||
| 36 | VideoMAE (no extra data, ViT-B, 16frame) | 70.8 | 92.4 | 87 | 180x6 | – | Paper | Code | 2022 | 9 of 13 ran · 4 unverified | report |
| 37 | AMD(ViT-S/16) | 70.2 | 92.5 | 22 | 57x6 | – | Paper | – | 2023 | no code linked | report |
| 38 | ILA (ViT-L/14) | 70.2 | 91.8 | – | Paper | Code | 2023 | 2 of 4 ran · 2 unverified | report | ||
| 39 | MorphMLP-B (IN-1K) | 70.1 | 92.8 | 68.5 | 197x3 | ✓ | Paper | Code | 2021 | 2 of 2 ran · 0 unverified | report |
| 40 | CoVeR(JFT-300M) | 69.8 | 91.9 | ✓ | Paper | – | 2021 | no code linked | report | ||
| 41 | TPS | 69.8 | – | Paper | Code | 2022 | 1 of 1 ran · 0 unverified | report | |||
| 42 | SIFA | 69.8 | – | Paper | Code | 2022 | linked, not harvested | report | |||
| 43 | Swin-B (IN-21K + Kinetics400 pretrain) | 69.6 | 92.7 | 89 | 321x3 | ✓ | Paper | Code | 2021 | 7 of 32 ran · 25 unverified | report |
| 44 | TDN ResNet101 (one clip, three crop, 8+16 ensemble, ImageNet pretrained, RGB only) | 69.6 | 92.2 | 198x3 | ✓ | Paper | Code | 2020 | 2 of 4 ran · 2 unverified | report | |
| 45 | MAR (75% mask, ViT-B, 16x4) | 69.5 | 91.9 | 94 | 41x6 | – | Paper | Code | 2022 | linked, not harvested | report |
| 46 | ORViT Mformer-L (ORViT blocks) | 69.5 | 91.5 | N/A | N/A | ✓ | Paper | Code | 2021 | 1 of 7 ran · 6 unverified | report |
| 47 | UniFormer-S (IN-1K + Kinetics600 pretrain) | 69.4 | 92.1 | 21.4 | 41.8x3 | ✓ | Paper | Code | 2021 | linked, not harvested | report |
| 48 | MML (ensemble) | 69.02 | 92.70 | ✓ | Paper | Code | 2020 | linked, not harvested | report | ||
| 49 | MViT-B-24, 32x3 | 68.7 | 91.5 | 53.2M | 236x3 | ✓ | Paper | Code | 2021 | 13 of 26 ran · 13 unverified | report |
| 50 | MTV-B | 68.5 | 90.4 | ✓ | Paper | Code | 2022 | linked, not harvested | report | ||
| 51 | MLP-3D | 68.5 | – | Paper | – | 2022 | no code linked | report | |||
| 52 | TDN ResNet101 (one clip, center crop, 8+16 ensemble, ImageNet pretrained, RGB only) | 68.2 | 91.6 | 198x1 | ✓ | Paper | Code | 2020 | 2 of 4 ran · 2 unverified | report | |
| 53 | MSMA (8+16frames) | 68.2 | – | Paper | – | 2023 | no code linked | report | |||
| 54 | Mformer-L | 68.1 | 91.2 | N/A | 1181x3 | ✓ | Paper | Code | 2021 | 1 of 2 ran · 1 unverified | report |
| 55 | VIMPAC | 68.1 | ✓ | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report | |||
| 56 | ORViT Mformer (ORViT blocks) | 67.9 | 90.5 | N/A | N/A | ✓ | Paper | Code | 2021 | 1 of 7 ran · 6 unverified | report |
| 57 | MViT-B, 32x3(Kinetics600 pretrain) | 67.8 | 91.3 | 36.6 | 170x3 | ✓ | Paper | Code | 2021 | 13 of 26 ran · 13 unverified | report |
| 58 | GC-TDN Ensemble (R50,8+16) | 67.8 | 91.2 | 27.4 | 110.1 | ✓ | Paper | Code | 2022 | 5 of 5 ran · 0 unverified | report |
| 59 | CT-Net Ensemble (R50, 8+12+16+24) | 67.8 | 91.1 | 83.8 | 280 | ✓ | Paper | Code | 2021 | 10 of 13 ran · 3 unverified | report |
| 60 | TCM (Ensemble) | 67.8 | – | Paper | Code | 2022 | linked, not harvested | report | |||
| 61 | SELFYNet-TSM-R50En (8+16 frames, ImageNet pretrained, 2 clips) | 67.7 | 91.1 | ✓ | Paper | Code | 2021 | 3 of 4 ran · 1 unverified | report | ||
| 62 | RSANet-R50 (8+16 frames, ImageNet pretrained, 2 clips | 67.7 | 91.1 | ✓ | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report | ||
| 63 | GTDNet | 67.6 | – | Paper | – | 2022 | no code linked | report | |||
| 64 | SELFYNet-TSM-R50En (8+16 frames, ImageNet pretrained, a single clip) | 67.4 | 91 | ✓ | Paper | Code | 2021 | 3 of 4 ran · 1 unverified | report | ||
| 65 | VoV3D-L (32frames, Kinetics pretrained, single) | 67.35 | 90.50 | 5.8M | 20.9x6 | ✓ | Paper | Code | 2020 | linked, not harvested | report |
| 66 | PLAR | 67.3 | 91 | – | Paper | – | 2023 | no code linked | report | ||
| 67 | RSANet-R50 (8+16 frames, ImageNet pretrained, a single clip) | 67.3 | 90.8 | ✓ | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report | ||
| 68 | X-Vit (x16) | 67.2 | 90.8 | N/A | 850x1 | ✓ | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report |
| 69 | TAda2D-En (ResNet-50, 8+16 frames) | 67.2 | 89.8 | ✓ | Paper | Code | 2021 | 0 of 1 ran · 1 unverified | report | ||
| 70 | Mformer-HR | 67.1 | 90.6 | N/A | 958.8x3 | ✓ | Paper | Code | 2021 | 1 of 2 ran · 1 unverified | report |
| 71 | TAdaConvNeXt-T | 67.1 | 90.4 | ✓ | Paper | Code | 2021 | 0 of 1 ran · 1 unverified | report | ||
| 72 | MoDS (8+16frames) | 67.1 | – | Paper | – | 2022 | no code linked | report | |||
| 73 | STPG (8+16frames) | 67.0 | – | Paper | – | 2022 | no code linked | report | |||
| 74 | MML (single) | 66.83 | 91.30 | ✓ | Paper | Code | 2020 | linked, not harvested | report | ||
| 75 | ILA (ViT-B/16) | 66.8 | 90.3 | – | Paper | Code | 2023 | 2 of 4 ran · 2 unverified | report | ||
| 76 | TSM (RGB + Flow) | 66.6 | 91.3 | ✓ | Paper | Code | 2018 | 6 of 16 ran · 10 unverified | report | ||
| 77 | MSNet-R50En (8+16 ensemble, ImageNet pretrained) | 66.6 | 90.6 | ✓ | Paper | Code | 2020 | 3 of 8 ran · 5 unverified | report | ||
| 78 | PAN ResNet101 (RGB only, no Flow) | 66.5 | 90.6 | ✓ | Paper | Code | 2020 | linked, not harvested | report | ||
| 79 | TSM+W3 (16 frames, RGB ResNet-50) | 66.5 | 90.4 | ✓ | Paper | – | 2020 | no code linked | report | ||
| 80 | Mformer | 66.5 | 90.1 | ✓ | Paper | Code | 2021 | 1 of 2 ran · 1 unverified | report | ||
| 81 | MVFNet-ResNet50 (center crop, 8+16 ensemble, ImageNet pretrained, RGB only) | 66.3 | ✓ | Paper | Code | 2020 | linked, not harvested | report | |||
| 82 | MViT-B, 16x4 | 66.2 | 90.2 | ✓ | Paper | Code | 2021 | 13 of 26 ran · 13 unverified | report | ||
| 83 | RSANet-R50 (16 frames, ImageNet pretrained, a single clip) | 66 | 89.8 | ✓ | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report | ||
| 84 | VoV3D-L (32frames, from scratch, single) | 65.8 | 89.5 | 5.8M | 20.9x6 | – | Paper | Code | 2020 | linked, not harvested | report |
| 85 | E3D-L | 65.7 | 89.8 | 18.3 | – | Paper | Code | 2023 | linked, not harvested | report | |
| 86 | SELFYNet-TSM-R50 (16 frames, ImageNet pretrained) | 65.7 | 89.8 | ✓ | Paper | Code | 2021 | 3 of 4 ran · 1 unverified | report | ||
| 87 | TAda2D (ResNet-50, 16 frames) | 65.6 | 89.2 | ✓ | Paper | Code | 2021 | 0 of 1 ran · 1 unverified | report | ||
| 88 | ViViT-L/16x2 Fact. encoder | 65.4 | 89.8 | ✓ | Paper | Code | 2021 | 14 of 21 ran · 7 unverified | report | ||
| 89 | VoV3D-M (32frames, Kinetics pretrained, single) | 65.24 | 89.48 | 3.3M | 11.5x6 | ✓ | Paper | Code | 2020 | linked, not harvested | report |
| 90 | bLVNet | 65.2 | ✓ | Paper | Code | 2019 | 0 of 6 ran · 6 unverified | report | |||
| 91 | DirecFormer | 64.94 | 87.9 | – | Paper | Code | 2022 | 4 of 8 ran · 4 unverified | report | ||
| 92 | RSANet-R50 (8 frames, ImageNet pretrained, a single clip) | 64.8 | 89.1 | ✓ | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report | ||
| 93 | MSNet-R50 (16 frames, ImageNet pretrained) | 64.7 | 89.4 | ✓ | Paper | Code | 2020 | 3 of 8 ran · 5 unverified | report | ||
| 94 | AK-Net | 64.3 | – | Paper | – | 2022 | no code linked | report | |||
| 95 | VoV3D-M (32frames, from scratch, single) | 64.2 | 88.8 | 3.3M | 11.5x6 | – | Paper | Code | 2020 | linked, not harvested | report |
| 96 | VoV3D-L (16frames, from scratch, single) | 64.1 | 88.6 | 5.8M | 9.3x6 | – | Paper | Code | 2020 | linked, not harvested | report |
| 97 | TAda2D (ResNet-50, 8 frames) | 64.0 | 88.0 | ✓ | Paper | Code | 2021 | 0 of 1 ran · 1 unverified | report | ||
| 98 | MoViNet-A2 | 63.5 | 89.0 | 4.8M | 10.3x1 | – | Paper | Code | 2021 | 8 of 13 ran · 5 unverified | report |
| 99 | VoV3D-M (16frames, from scratch, single) | 63.2 | 88.2 | 3.3M | 5.7x6 | – | Paper | Code | 2020 | linked, not harvested | report |
| 100 | MSNet-R50 (8 frames, ImageNet pretrained) | 63 | 88.4 | ✓ | Paper | Code | 2020 | 3 of 8 ran · 5 unverified | report | ||
| 101 | MoViNet-A1 | 62.7 | 89.0 | 4.6M | 6.0x1 | – | Paper | Code | 2021 | 8 of 13 ran · 5 unverified | report |
| 102 | OmniVL | 62.5 | 86.2 | – | Paper | – | 2022 | no code linked | report | ||
| 103 | TimeSformer-HR | 62.5 | ✓ | Paper | Code | 2021 | 35 of 43 ran · 8 unverified | report | |||
| 104 | TimeSformer-L | 62.3 | ✓ | Paper | Code | 2021 | 35 of 43 ran · 8 unverified | report | |||
| 105 | TRG (ResNet-50) | 62.2 | 90.3 | – | Paper | – | 2019 | no code linked | report | ||
| 106 | TPN (TSM-50) | 62.0 | – | Paper | Code | 2020 | 0 of 3 ran · 3 unverified | report | |||
| 107 | Multigrid | 61.7 | ✓ | Paper | Code | 2019 | 2 of 10 ran · 8 unverified | report | |||
| 108 | SlowFast | 61.7 | ✓ | Paper | Code | 2018 | 0 of 10 ran · 10 unverified | report | |||
| 109 | TRG (Inception-V3) | 61.3 | 91.4 | – | Paper | – | 2019 | no code linked | report | ||
| 110 | MoViNet-A0 | 61.3 | 88.2 | 3.1M | 2.7x1 | – | Paper | Code | 2021 | 8 of 13 ran · 5 unverified | report |
| 111 | CCS + two-stream + TRN | 61.2 | 89.3 | – | Paper | – | 2019 | no code linked | report | ||
| 112 | VidTr-L | 60.2 | – | Paper | – | 2021 | no code linked | report | |||
| 113 | TimeSformer | 59.5 | ✓ | Paper | Code | 2021 | 35 of 43 ran · 8 unverified | report | |||
| 114 | SVT | 59.2 | – | Paper | Code | 2021 | 0 of 1 ran · 1 unverified | report | |||
| 115 | TAM (5-shot) | 52.3 | – | Paper | – | 2019 | no code linked | report | |||
| 116 | model3D_1 with left-right augmentation and fps jitter | 51.33 | 80.46 | – | Paper | Code | 2017 | linked, not harvested | report | ||
| 117 | Prob-Distill | 49.9 | 79.1 | – | Paper | – | 2019 | no code linked | report | ||
| 118 | STM + TRNMultiscale | 47.73 | – | Paper | Code | 2019 | linked, not harvested | report | |||
| 119 | InternVideo2-6B | 1 | 12 | 2131 | 13321 | ✓ | Paper | Code | 2024 | linked, not harvested | report |
| 120 | MViTv2-B (IN-21K + Kinetics400 pretrain) | 93.4 | 51.1 | – | Paper | Code | 2021 | linked, not harvested | report | ||
| 121 | RSANet-R50 (8+16 frames, ImageNet pretrained, 2 clips) | 91.1 | – | Paper | Code | 2021 | 1 of 1 ran · 0 unverified | report | |||
| 122 | MoViNet-A3 | 5.3M | 23.7x1 | – | Paper | Code | 2021 | 8 of 13 ran · 5 unverified | report | ||
| 123 | MViT-L (IN-21K + Kinetics400 pretrain) | 2828x3 | ✓ | Paper | Code | 2021 | linked, not harvested | report |
All 123 rows shown. 123 link to a paper page on this site; 70 are marked as using additional training data in the archive. No GitHub stars are tracked; "Code" is the first repository the archive lists for the row. The archive carries no row tags, review links or community-submitted rows for this table; none are shown. archive 2025-07-28
Syntology Ran reads "N of M ran · U unverified": of the M code samples Syntology harvested from repositories linked to that row's paper (joined by arXiv id), N executed on a synthesized input and the other U = M−N are unverified (harvested, no recorded run). It counts code from repositories linked to that row's paper, not this result: the row's number was not reproduced and nothing here is a correctness claim. The other cell texts mean no graph line for the row: "linked, not harvested" (the archive links code, Syntology has not harvested it), "no code linked" (no code link in the archive), "not matched" (the row's paper URL matched no paper on this site). 63 rows have a graph line, from 37 distinct papers; 54 rows (31 papers) have at least one sample that ran. Counting each paper once: Syntology ran 158 of 314 samples; 156 unverified. Separately, 84 of those 314 are pointer-only (licence): the site points at that code rather than redistributing it, a licence property recorded for ran and unverified samples alike; each cell's tooltip carries the row's own pointer-only count. Read from the graph 2026-09-24. Per-sample status is on the paper page.
Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections