Update README.md
Browse files
README.md
CHANGED
|
@@ -1,117 +1,180 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: cc-by-4.0
|
| 3 |
-
language:
|
| 4 |
-
- en
|
| 5 |
-
metrics:
|
| 6 |
-
- roc_auc
|
| 7 |
-
- matthews_correlation
|
| 8 |
-
pipeline_tag: video-classification
|
| 9 |
-
---
|
| 10 |
-
# Simplifying Traffic Anomaly Detection with Video Foundation Models
|
| 11 |
-
|
| 12 |
-
Svetlana Orlova, Tommie Kerssies,
|
| 13 |
-
Eindhoven University of Technology
|
| 14 |
-
|
| 15 |
-
[](https://arxiv.org/abs/2507.09338)
|
| 16 |
-
[](https://huggingface.co/tue-mps/simple-tad/tree/main/models)
|
| 17 |
-
[](https://github.com/tue-mps/simple-tad)
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
###
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
|
| 55 |
-
|
|
| 56 |
-
|
|
| 57 |
-
|
|
| 58 |
-
|
|
| 59 |
-
|
|
| 60 |
-
|
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
|
| 66 |
-
|
|
| 67 |
-
| VideoMAE | ViT-
|
| 68 |
-
| VideoMAE | ViT-
|
| 69 |
-
|
|
| 70 |
-
| VideoMAE2 | ViT-
|
| 71 |
-
|
|
| 72 |
-
|
|
| 73 |
-
|
|
| 74 |
-
|
|
| 75 |
-
| **DAPT-** VideoMAE | ViT-
|
| 76 |
-
| **DAPT-** VideoMAE | ViT-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
#
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
[
|
| 89 |
-
[
|
| 90 |
-
[
|
| 91 |
-
[
|
| 92 |
-
[
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
##
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
metrics:
|
| 6 |
+
- roc_auc
|
| 7 |
+
- matthews_correlation
|
| 8 |
+
pipeline_tag: video-classification
|
| 9 |
+
---
|
| 10 |
+
# Simplifying Traffic Anomaly Detection with Video Foundation Models
|
| 11 |
+
|
| 12 |
+
Svetlana Orlova, Tommie Kerssies, Brunó B. Englert, Gijs Dubbelman \
|
| 13 |
+
Eindhoven University of Technology
|
| 14 |
+
|
| 15 |
+
[](https://arxiv.org/abs/2507.09338)
|
| 16 |
+
[](https://huggingface.co/tue-mps/simple-tad/tree/main/models)
|
| 17 |
+
[](https://github.com/tue-mps/simple-tad)
|
| 18 |
+
|
| 19 |
+
<table>
|
| 20 |
+
<tr>
|
| 21 |
+
<td>
|
| 22 |
+
<a href="https://youtu.be/hY2hUlTNhCU" target="_blank">
|
| 23 |
+
<img src="illustrations/videos/PYL3JcSsS6o_004036_small-dapt.gif" width="100%">
|
| 24 |
+
</a>
|
| 25 |
+
</td>
|
| 26 |
+
<td>
|
| 27 |
+
<a href="https://youtu.be/tKe2nTIHf9k" target="_blank">
|
| 28 |
+
<img src="illustrations/videos/Sihe6aeyLHg_000602_small-dapt.gif" width="100%">
|
| 29 |
+
</a>
|
| 30 |
+
</td>
|
| 31 |
+
</tr>
|
| 32 |
+
</table>
|
| 33 |
+
|
| 34 |
+
Recent methods for ego-centric Traffic Anomaly Detection (TAD) often rely on complex multi-stage or multi-representation fusion architectures, yet it remains unclear whether such complexity is necessary. Recent findings in visual perception suggest that foundation models, enabled by advanced pre-training, allow simple yet flexible architectures to outperform specialized designs. Therefore, in this work, we investigate an architecturally simple encoder-only approach using plain Video Vision Transformers (Video ViTs) and study how pre-training enables strong TAD performance. We find that: (i) advanced pre-training enables simple encoder-only models to match or even surpass the performance of specialized state-of-the-art TAD methods, while also being significantly more efficient; (ii) although weakly- and fully-supervised pre-training are advantageous on standard benchmarks, we find them less effective for TAD. Instead, self-supervised Masked Video Modeling (MVM) provides the strongest signal; and (iii) Domain-Adaptive Pre-Training (DAPT) on unlabeled driving videos further improves downstream performance, without requiring anomalous examples. Our findings highlight the importance of pre-training and show that effective, efficient, and scalable TAD models can be built with minimal architectural complexity.
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
### ✨ DoTA and DADA-2000 results
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
Video ViT-based encoder-only models set a new state of the art
|
| 43 |
+
on both datasets, while being significantly more efficient than top-performing specialized methods. FPS measured using NVIDIA A100
|
| 44 |
+
MIG, 2 1 GPU. † From prior work. ‡ Optimistic estimates using publicly available components of the model. “A→B”: trained on A, tested
|
| 45 |
+
on B; D2K: DADA-2000.
|
| 46 |
+
|
| 47 |
+
## 🧩 Code
|
| 48 |
+
|
| 49 |
+
Check out our GitHub repo: **[simple-tad](https://github.com/tue-mps/simple-tad)**
|
| 50 |
+
|
| 51 |
+
## 📍Model Zoo
|
| 52 |
+
|
| 53 |
+
### DAPT (adapted) models
|
| 54 |
+
| Method | Backbone | Initialized with | DAPT epochs | DAPT data | Checkpoint |
|
| 55 |
+
| :------: |:--------:|:---------------:|:-----------:|:------------------:|:----------:|
|
| 56 |
+
| VideoMAE | ViT-S | [Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400) | 12 | Kinetics-700 | [simpletad_dapt-k700_videomae-s_ep12.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/DAPT/simpletad_dapt-k700_videomae-s_ep12.pth) |
|
| 57 |
+
| VideoMAE | ViT-S | [Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400) | 12 | BDD100K | [simpletad_dapt-onlybdd_videomae-s_ep12.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/DAPT/simpletad_dapt-onlybdd_videomae-s_ep12.pth) |
|
| 58 |
+
| VideoMAE | ViT-S | [Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400) | 12 | BDD100K + CAP-DATA | [simpletad_dapt_videomae-s_ep12.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/DAPT/simpletad_dapt_videomae-s_ep12.pth) |
|
| 59 |
+
| VideoMAE | ViT-B | [Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400) | 12 | BDD100K + CAP-DATA | [simpletad_dapt_videomae-b_ep12.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/DAPT/simpletad_dapt_videomae-b_ep12.pth) |
|
| 60 |
+
| VideoMAE | ViT-L | [Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400) | 12 | BDD100K + CAP-DATA | [simpletad_dapt_videomae-l_ep12.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/DAPT/simpletad_dapt_videomae-l_ep12.pth) |
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
### Fine-tuned on DoTA
|
| 64 |
+
| Method | Backbone | Initialized with | Best AUC<sub>ROC</sub> checkpoint | Best AUC<sub>MCC</sub> checkpoint | AUC<sub>ROC</sub> | AUC<sub>MCC</sub> |
|
| 65 |
+
|:---------:|:--------:|:--------------------------------:|:-----------------------:|:-----------------------:|:-------:|---------|
|
| 66 |
+
| VideoMAE | ViT-S | VideoMAE ([Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dota_vm1-s_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm1-s_auroc.pth) | [simpletad_ft-dota_vm1-s_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm1-s_aumcc.pth) | 83.7 | 46.9 |
|
| 67 |
+
| VideoMAE | ViT-B | VideoMAE ([Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dota_vm1-b-1600_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm1-b-1600_auroc.pth) | [simpletad_ft-dota_vm1-b-1600_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm1-b-1600_aumcc.pth) | 86.3 | 54.8 |
|
| 68 |
+
| VideoMAE | ViT-L | VideoMAE ([Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dota_vm1-l_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm1-l_auroc.pth) | [simpletad_ft-dota_vm1-l_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm1-l_aumcc.pth) | 88.2 | 58.7 |
|
| 69 |
+
| VideoMAE2 | ViT-S | VideoMAE2 ([vit_s_k710_dl_from_giant.pth](https://github.com/OpenGVLab/VideoMAEv2/blob/master/docs/MODEL_ZOO.md#distillation)) | [simpletad_ft-dota_vm2-s_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm2-s_auroc.pth) | [simpletad_ft-dota_vm2-s_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm2-s_aumcc.pth) | 86.0 | 54.1 |
|
| 70 |
+
| VideoMAE2 | ViT-B | VideoMAE2 ([vit_b_k710_dl_from_giant.pth](https://github.com/OpenGVLab/VideoMAEv2/blob/master/docs/MODEL_ZOO.md#distillation)) | [simpletad_ft-dota_vm2-b_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm2-b_auroc.pth) | [simpletad_ft-dota_vm2-b_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_vm2-b_aumcc.pth) | 86.9 | 55.4 |
|
| 71 |
+
| MVD_fromL | ViT-S | MVD ([Kinetics-400 Teacher ViT-L](https://github.com/ruiwang2021/mvd/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dota_mvd-s-fromL_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_mvd-s-fromL_auroc.pth) | [simpletad_ft-dota_mvd-s-fromL_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_mvd-s-fromL_aumcc.pth) | 85.3 | 53.8 |
|
| 72 |
+
| MVD_fromB | ViT-B | MVD ([Kinetics-400 Teacher ViT-B](https://github.com/ruiwang2021/mvd/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dota_mvd-b-fromB_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_mvd-b-fromB_auroc.pth) | [simpletad_ft-dota_mvd-b-fromB_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_mvd-b-fromB_aumcc.pth) | 86.1 | 54.7 |
|
| 73 |
+
| MVD_fromL | ViT-L | MVD ([Kinetics-400 Teacher ViT-L](https://github.com/ruiwang2021/mvd/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dota_mvd-l-fromL_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_mvd-l-fromL_auroc.pth) | [simpletad_ft-dota_mvd-l-fromL_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_mvd-l-fromL_aumcc.pth) | 87.2 | 58.1 |
|
| 74 |
+
| **DAPT-** VideoMAE | ViT-S | DAPT (BDD100K + CAP-DATA) | [simpletad_ft-dota_dapt-vm1-s_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_dapt-vm1-s_auroc.pth) | [simpletad_ft-dota_dapt-vm1-s_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_dapt-vm1-s_aumcc.pth) | 86.4 | 54.0 |
|
| 75 |
+
| **DAPT-** VideoMAE | ViT-B | DAPT (BDD100K + CAP-DATA) | [simpletad_ft-dota_dapt-vm1-b_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_dapt-vm1-b_auroc.pth) | [simpletad_ft-dota_dapt-vm1-b_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_dapt-vm1-b_aumcc.pth) | 87.9 | 57.5 |
|
| 76 |
+
| **DAPT-** VideoMAE | ViT-L | DAPT (BDD100K + CAP-DATA) | [simpletad_ft-dota_dapt-vm1-l_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_dapt-vm1-l_auroc.pth) | [simpletad_ft-dota_dapt-vm1-l_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_DoTA/simpletad_ft-dota_dapt-vm1-l_aumcc.pth) | 88.4 | 58.9 |
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
### Fine-tuned on DADA-2000
|
| 80 |
+
| Method | Backbone | Initialized with | Best AUC<sub>ROC</sub> checkpoint | Best AUC<sub>MCC</sub> checkpoint | AUC<sub>ROC</sub> | AUC<sub>MCC</sub> |
|
| 81 |
+
|:---------:|:--------:|:--------------------------------:|:-----------------------:|:-----------------------:|:-------:|---------|
|
| 82 |
+
| VideoMAE | ViT-S | VideoMAE ([Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dada_vm1-s_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm1-s_auroc.pth) | [simpletad_ft-dada_vm1-s_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm1-s_aumcc.pth) | 83.0 | 48.2 |
|
| 83 |
+
| VideoMAE | ViT-B | VideoMAE ([Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dada_vm1-b-1600_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm1-b-1600_auroc.pth) | [simpletad_ft-dada_vm1-b-1600_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm1-b-1600_aumcc.pth) | 85.4 | 52.2 |
|
| 84 |
+
| VideoMAE | ViT-L | VideoMAE ([Kinetics-400 1600 ep](https://github.com/MCG-NJU/VideoMAE/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dada_vm1-l_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm1-l_auroc.pth) | [simpletad_ft-dada_vm1-l_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm1-l_aumcc.pth) | 87.2 | 55.4 |
|
| 85 |
+
| VideoMAE2 | ViT-S | VideoMAE2 ([vit_s_k710_dl_from_giant.pth](https://github.com/OpenGVLab/VideoMAEv2/blob/master/docs/MODEL_ZOO.md#distillation)) | [simpletad_ft-dada_vm2-s_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm2-s_auroc.pth) | [simpletad_ft-dada_vm2-s_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm2-s_aumcc.pth) | 84.8 | 50.3 |
|
| 86 |
+
| VideoMAE2 | ViT-B | VideoMAE2 ([vit_b_k710_dl_from_giant.pth](https://github.com/OpenGVLab/VideoMAEv2/blob/master/docs/MODEL_ZOO.md#distillation)) | [simpletad_ft-dada_vm2-b_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm2-b_auroc.pth) | [simpletad_ft-dada_vm2-b_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_vm2-b_aumcc.pth) | 86.3 | 53.3 |
|
| 87 |
+
| MVD_fromL | ViT-S | MVD ([Kinetics-400 Teacher ViT-L](https://github.com/ruiwang2021/mvd/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dada_mvd-s-fromL_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_mvd-s-fromL_auroc.pth) | [simpletad_ft-dada_mvd-s-fromL_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_mvd-s-fromL_aumcc.pth) | 82.2 | 50.2 |
|
| 88 |
+
| MVD_fromB | ViT-B | MVD ([Kinetics-400 Teacher ViT-B](https://github.com/ruiwang2021/mvd/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dada_mvd-b-fromB_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_mvd-b-fromB_auroc.pth) | [simpletad_ft-dada_mvd-b-fromB_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_mvd-b-fromB_aumcc.pth) | 84.7 | 50.9 |
|
| 89 |
+
| MVD_fromL | ViT-L | MVD ([Kinetics-400 Teacher ViT-L](https://github.com/ruiwang2021/mvd/blob/main/MODEL_ZOO.md#kinetics-400)) | [simpletad_ft-dada_mvd-l-fromL_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_mvd-l-fromL_auroc.pth) | [simpletad_ft-dada_mvd-l-fromL_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_mvd-l-fromL_aumcc.pth) | 86.1 | 53.7 |
|
| 90 |
+
| **DAPT-** VideoMAE | ViT-S | DAPT (BDD100K + CAP-DATA) | [simpletad_ft-dada_dapt-vm1-s_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_dapt-vm1-s_auroc.pth) | [simpletad_ft-dada_dapt-vm1-s_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_dapt-vm1-s_aumcc.pth) | 85.6 | 52.0 |
|
| 91 |
+
| **DAPT-** VideoMAE | ViT-B | DAPT (BDD100K + CAP-DATA) | [simpletad_ft-dada_dapt-vm1-b_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_dapt-vm1-b_auroc.pth) | [simpletad_ft-dada_dapt-vm1-b_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_dapt-vm1-b_aumcc.pth) | 87.6 | 55.2 |
|
| 92 |
+
| **DAPT-** VideoMAE | ViT-L | DAPT (BDD100K + CAP-DATA) | [simpletad_ft-dada_dapt-vm1-l_auroc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_dapt-vm1-l_auroc.pth) | [simpletad_ft-dada_dapt-vm1-l_aumcc.pth](https://huggingface.co/tue-mps/simple-tad/resolve/main/models/Finetune_D2K/simpletad_ft-dada_dapt-vm1-l_aumcc.pth) | 88.5 | 56.8 |
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
## ☎️ Contact
|
| 96 |
+
|
| 97 |
+
Svetlana Orlova: s.orlova@tue.nl, orsveri@gmail.com
|
| 98 |
+
|
| 99 |
+
## 👍 Acknowledgements
|
| 100 |
+
|
| 101 |
+
Our code is mainly based on the [VideoMAE](https://github.com/MCG-NJU/VideoMAE) codebase.
|
| 102 |
+
With Video ViTs that have identical architecture, we only used their weights:
|
| 103 |
+
[ViViT](https://github.com/google-research/scenic/blob/main/scenic/projects/vivit/README.md),
|
| 104 |
+
[VideoMAE2](https://github.com/OpenGVLab/VideoMAEv2),
|
| 105 |
+
[SMILE](https://github.com/fmthoker/SMILE),
|
| 106 |
+
[SIGMA](https://github.com/QUVA-Lab/SIGMA/),
|
| 107 |
+
[MME](https://github.com/XinyuSun/MME),
|
| 108 |
+
[MGMAE](https://github.com/MCG-NJU/MGMAE). \
|
| 109 |
+
We used fragments of original implementations of
|
| 110 |
+
[MVD](https://github.com/ruiwang2021/mvd),
|
| 111 |
+
[InternVideo2](https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo2/single_modality),
|
| 112 |
+
and [UMT](https://github.com/OpenGVLab/unmasked_teacher/tree/main/single_modality) to integrate these models with our codebase.
|
| 113 |
+
|
| 114 |
+
## ✏️ Citation
|
| 115 |
+
|
| 116 |
+
If you think this project is helpful, please feel free to like us ❤️ and cite our paper:
|
| 117 |
+
|
| 118 |
+
```
|
| 119 |
+
@inproceedings{orlova2025simplifying,
|
| 120 |
+
title={Simplifying Traffic Anomaly Detection with Video Foundation Models},
|
| 121 |
+
author={Orlova, Svetlana and Kerssies, Tommie and Englert, Brun{\'o} B and Dubbelman, Gijs},
|
| 122 |
+
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
|
| 123 |
+
year={2025}
|
| 124 |
+
}
|
| 125 |
+
|
| 126 |
+
@article{orlova2025simplifying,
|
| 127 |
+
title={Simplifying Traffic Anomaly Detection with Video Foundation Models},
|
| 128 |
+
author={Orlova, Svetlana and Kerssies, Tommie and Englert, Brun{\'o} B and Dubbelman, Gijs},
|
| 129 |
+
journal={arXiv preprint arXiv:2507.09338},
|
| 130 |
+
year={2025}
|
| 131 |
+
}
|
| 132 |
+
```
|
| 133 |
+
|
| 134 |
+
##
|
| 135 |
+
|
| 136 |
+
<table>
|
| 137 |
+
<tr>
|
| 138 |
+
<td>
|
| 139 |
+
<a href="https://youtu.be/vPBKj9SF9yg" target="_blank">
|
| 140 |
+
<img src="illustrations/videos/0RJPQ_97dcs_004503_large-dapt.gif" width="100%">
|
| 141 |
+
</a>
|
| 142 |
+
</td>
|
| 143 |
+
<td>
|
| 144 |
+
<a href="https://youtu.be/vGaYPZEuv5k" target="_blank">
|
| 145 |
+
<img src="illustrations/videos/y4Evv5By6sg_004171_large-dapt.gif" width="100%">
|
| 146 |
+
</a>
|
| 147 |
+
</td>
|
| 148 |
+
</tr>
|
| 149 |
+
</table>
|
| 150 |
+
|
| 151 |
+
<table>
|
| 152 |
+
<tr>
|
| 153 |
+
<td>
|
| 154 |
+
<a href="https://youtu.be/7rH0QP18zsk" target="_blank">
|
| 155 |
+
<img src="illustrations/videos/PEwiwzyTjX0_000589large-dapt.gif" width="100%">
|
| 156 |
+
</a>
|
| 157 |
+
</td>
|
| 158 |
+
<td>
|
| 159 |
+
<a href="https://youtu.be/5ZNYwDGmOZI" target="_blank">
|
| 160 |
+
<img src="illustrations/videos/RASKiMoxhOE_000246_large-dapt.gif" width="100%">
|
| 161 |
+
</a>
|
| 162 |
+
</td>
|
| 163 |
+
</tr>
|
| 164 |
+
</table>
|
| 165 |
+
|
| 166 |
+
<table>
|
| 167 |
+
<tr>
|
| 168 |
+
<td>
|
| 169 |
+
<a href="https://youtu.be/X7Ij1sc4yCE" target="_blank">
|
| 170 |
+
<img src="illustrations/videos/T7TkJVmGyts_001011_large-dapt.gif" width="100%">
|
| 171 |
+
</a>
|
| 172 |
+
</td>
|
| 173 |
+
<td>
|
| 174 |
+
<a href="https://youtu.be/S5m2ooY6CGc" target="_blank">
|
| 175 |
+
<img src="illustrations/videos/TNZv-NBcV5U_000066large-dapt.gif" width="100%">
|
| 176 |
+
</a>
|
| 177 |
+
</td>
|
| 178 |
+
</tr>
|
| 179 |
+
</table>
|
| 180 |
+
|