Organizations: Visual and Spatial AI Lab, VCCM Section, Texas A&M University, College Station, Texas, USA · Department of ECEN, Texas A&M University, College Station, Texas, USA · College of PVFA, Texas A&M University, College Station, Texas, USA · Department of CSCE, Texas A&M University, College Station, Texas, USA
We investigate how frozen pretrained visual and depth representations support dense optical flow when coupled with a supervised matching network. Building on TTSFlow, this paper studies Dense Prediction Transformer (DPT) feature decoding and one coarse-to-fine refinement stage. Our method fuses features from frozen DINOv2 and Depth Anything V2 models and estimates optical flow through a trainable transformer-based matching pipeline. After optical-flow training on FlyingChairs and FlyingThings3D, it achieves 2.37 endpoint error (EPE) on the Sintel training split, Final pass, compared with 2.52 for the FlowSeek (T) configuration reported here. Our RoCo-Spring 2026 entry, RoCo-29.VSAI, obtains a robustness-balanced score of 1.1038, Spring EPE of 1.411, and RobustSpring proxy EPE of 4.739, with a lower balanced score than seven of the twelve listed reference baselines. These results characterize the complete pipeline, which reuses frozen foundation backbones and performs one refinement stage. Code is available at: https://github.com/Visual-and-Spatial-AI-Lab/FFlow
Figures & tables
Figure 1: Overall pipeline, including refinement. Each feature branch is applied to both RGB frames. The TTSFlow feature-fusion design [ 3 ] is extended with visual DPT decoding and two feature scales. Matching, propagation, refinement, and convex upsampling follow GMFlow [ 55 ] . The transformer is shared across scales. Fusion and projection weights are shared. Frozen labels apply to the pretrained foundation models. The visual DPT gradient status in the released code is documented in Appendix G.2 .
Method
#refine
Things (val)
Sintel (train, clean)
Sintel (train, final)
EPE
EPE
s 0-10
s 10-40
s 40+
EPE
s 0-10
s 10-40
s 40+
RAFT [ 47 ]
32
4.25
1.43
0.33
1.54
9.04
2.71
0.51
2.98
17.62
GMFlow [ 55 ]
0
3.48
1.50
0.46
1.77
8.26
2.96
0.72
3.45
17.70
GMFlow [ 55 ]
1
2.80
1.08
0.30
1.25
6.26
2.48
0.51
2.81
15.76
SEA-RAFT (S) [ 51 ]
4
-
1.27
-
-
-
4.32
-
-
-
FlowSeek (T) [ 34 ]
4
3.94
1.12
0.23
1.27
7.12
2.52
0.45
2.78
16.60
Table 1: Generalization performance after training on Chairs and Things train sets. The 2.37 EPE is on the Sintel training split, Final pass, and is the lowest among these listed configurations. The TTSFlow row is reproduced from [ 3 ] ; refinement counts are not runtime measurements.
Method
Extra flow data
#refine
Sintel (train)
KITTI (train)
Clean EPE
Final EPE
EPE
F1-all (%)
RAFT [ 47 ]
–
32
0.768
1.217
0.63
1.47
GMFlow [ 55 ]
–
0
0.947
1.276
2.06
7.57
GMFlow [ 55 ]
–
1
0.762
1.110
1.36
5.17
SEA-RAFT (S) [ 51 ]
TartanAir
4
0.546
0.782
0.93
2.65
FlowSeek (T) [ 34 ]
TartanAir
4
0.710
1.280
1.26
3.90
Table 2: Performance on the Sintel and KITTI training sets. Sintel results are obtained after training on Chairs, Things, and the mixed TSKH datasets. Training configurations differ; the comparison does not isolate architecture or training-scale effects. Extra flow data denotes additional optical-flow pretraining, not foundation-model pretraining. The TTSFlow row is reproduced from [ 3 ] . #refine denotes the number of refinement steps. Lower values indicate better performance for all metrics.
Method
RbS-Score
Spring EPE
Proxy EPE
Δ
SEA-RAFT [ 51 ]
0.4584
0.363
3.323
2.960
MS-RAFT+ [ 14 ]
0.6774
0.643
4.263
3.620
FlowFormer [ 12 ]
0.7368
0.723
4.493
3.770
GMFlow [ 55 ]
0.8015
0.945
3.925
2.980
GMA [ 15 ]
0.8704
0.914
4.944
4.030
VSAI (RoCo-29, Method 1)
1.1038
1.411
4.739
3.328
Table 3: RoCo-Spring optical flow result. The reported columns compare the submitted entry with challenge reference methods. Δ=ER,proxy−ES is shown for interpretation. All metrics are lower-is-better; EPE and disagreement are in pixels. Reference baselines are unranked challenge references, not competing team ranks.
Figure 2: Qualitative results on Spring Benchmark Dataset [ 41 ] .
Method
Things (val, clean)
Sintel (train, clean)
Sintel (train, final)
EPE
EPE
s 0-10
s 10-40
s 40+
EPE
s 0-10
s 10-40
s 40+
(w/o) Fusion
3.52
1.575
0.45
2.01
8.53
3.12
0.685
3.57
19.37
(w/) Fusion
3.02
1.46
0.39
2.05
7.74
2.81
0.64
3.36
16.99
Table 4: Fusion ablation reproduced from TTSFlow [ 3 ] , Table 4, with 800k Things-stage updates following 200k Chairs updates, and no DPT or refinement. This is prior evidence for the base fusion design, not a new ablation of the submitted configuration.
Method
Sintel (train, clean)
Sintel (train, final)
EPE
s0−10
s10−40
s40+
EPE
s0−10
s10−40
s40+
without DPT
1.33
0.40
1.61
7.26
2.73
0.62
3.16
16.69
with DPT
1.26
0.34
1.52
7.25
2.37
0.55
2.92
14.05
Table 5: DPT comparison on the Sintel training split, Clean and Final passes, after 360k Things-stage updates following Chairs training. Both configurations use one refinement stage.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Subset
EPE
Subset
EPE
Overall
1.411
Non-rigid
3.804
Low-detail
1.374
Not sky
1.406
High-detail
7.193
Sky
1.489
Matched
1.288
s0−10
0.898
Unmatched
6.500
s10−40
1.340
Rigid
1.094
s40+
5.887
Appendix
Table 6: Clean Spring EPE by evaluation subset. Values are in pixels.
Corruption
Δc
Corruption
Δc
Brightness
0.308
JPEG compression
1.671
Contrast
3.213
Motion blur
1.538
Defocus blur
0.592
Pixelate
0.756
Elastic transform
1.189
Rain
13.034
Fog
1.382
Saturate
0.415
Frost
8.579
Shot noise
1.980
Appendix
Table 7: RobustSpring disagreement for all twenty corruptions. Δc is in pixels. Lower values indicate less change in predicted flow. These are not corrupted-input ground-truth EPE values.
Figure 3: Sintel qualitative comparison. Two examples from the Sintel benchmark [ 1 ] . The circles are annotations retained from the reported experiment. The visual comparisons illustrate local differences in predicted motion, without establishing a quantitative occlusion or boundary advantage.
Figure 4: KITTI qualitative comparison. Two examples from the KITTI training split [ 10 ] . These examples belong to the target-adapted evaluation and should not be interpreted as held-out test results.
Figure 5: Predicted flow for the component-removal configurations in Table 8 , without DPT or refinement. Colors encode flow, not ground-truth error.
Method
Sintel (train, clean)
Sintel (train, final)
EPE
s 0-10
s 10-40
s 40+
EPE
s 0-10
s 10-40
s 40+
Ours (Depth)
122.87
103.32
152.01
197.83
120.12
99.42
152.38
196.38
Ours (DINOv2)
36.15
26.49
47.55
79.76
36.42
25.33
49.20
87.06
Ours (Depth + DINOv2)
1.46
0.39
2.05
7.74
2.81
0.64
3.36
16.99
Appendix
Table 8: Ablation of DINOv2 and depth features. Results on the Sintel training set for the reported component-removal configurations of a Chairs+Things model, without DPT or refinement. These are not a capacity-matched comparison of independently trained single-branch models.
(w/o) Depth
(w/o) DINO
Depth+DINO
EPE
1.77
1.07
0.87
Appendix
Table 9: Three feature configurations after 100k FlyingChairs updates, without DPT or refinement. The 1.77 and 0.87 entries are retained from TTSFlow [ 3 ] . The historical evaluation split is unspecified.
DepthAnythingV2 (Small)
DepthAnythingV2 (Base)
EPE (FlyingChairs)
0.995
0.87
Appendix
Table 10: DepthAnythingV2 backbone comparison. Results on the FlyingChairs dataset after training for 100k iterations.
DINOv2 (Small)
DINOv2 (Base)
EPE (FlyingChairs)
1.091
1.056
Appendix
Table 11: DINOv2 backbone comparison. Results on the FlyingChairs dataset after training for 50k iterations.
RAFT
GMFlow
GMFlow
FlowSeek[T]
SEA-RAFT[S]
Ours
#Refine
32
0
1
4
4
1
Time (ms)
229
194
295
144
63
362
Appendix
Table 12: Inference time comparison of different optical flow methods.
Method
#refine
KITTI EPE
KITTI F1-all
Ours (without DPT)
1
1.337
4.74
Ours (with DPT)
1
1.29
4.29
Appendix
Table 13: Performance on the KITTI training set after 140k KITTI fine-tuning updates, following Chairs, Things, and mixed TSKH training. Both configurations use one refinement stage; this is separate from the 200k setting in Table 2 .
Method
#refine
Things (val)
Sintel (train, clean)
Sintel (train, final)
EPE
EPE
s 0-10
s 10-40
s 40+
EPE
s 0-10
s 10-40
s 40+
RAFT [ 47 ]
32
4.25
1.43
0.33
1.54
9.04
2.71
0.51
2.98
17.62
GMFlow [ 55 ]
0
3.48
1.50
0.46
1.77
8.26
2.96
0.72
3.45
17.70
GMFlow [ 55 ]
1
2.80
1.08
0.30
1.25
6.26
2.48
0.51
2.81
15.76
SEA-RAFT (S) [ 51 ]
4
-
1.27
-
-
-
4.32
-
-
-
FlowSeek (T) [ 34 ]
4
3.94
1.12
0.23
1.27
7.12
2.52
0.45
2.78
16.60
Appendix
Table 14: Generalization after Chairs training and the indicated Things-stage training duration. The 560k checkpoint improves Things and Sintel Clean EPE relative to the 360k checkpoint, while Sintel Final EPE increases from 2.37 to 2.42.
Module
Parameters
Status in released flow path
DINOv2-S encoder
22,056,576
Frozen
Added visual DPT head
4,124,576
No flow gradient; flag true
Depth Anything V2-B encoder
86,580,480
Frozen
Depth Anything V2 DPT decoder
10,890,305
Frozen
Depth downsampling
959,040
Gradient-connected
Quarter-scale depth projection
32,896
Gradient-connected
Appendix
Table 15: Source-derived parameter inventory of the released two-scale configuration. Counts include biases and embedding/normalization parameters; shared modules are counted once. “No flow gradient” describes the released computation, not a reconstruction of historical training. No experimental checkpoint was loaded to obtain these counts.
Stage
Updates
Global batch
Crop (H×W)
Peak learning rate
FlyingChairs
200k
8
384×512
4×10−4
FlyingThings3D
360k
8
384×768
2×10−4
TSKH mixture
320k
4
320×896
2×10−4
KITTI
200k
4
320×1152 (code)
2×10−4
Spring
400k
4
544×960
1×10−4
Appendix
Table 16: Reported stage durations and released launcher settings. Durations denote updates within each stage, rather than cumulative updates over all stages. The launcher uses step limits one larger than the displayed round values and names checkpoints at the round values. The KITTI crop discrepancy is recorded explicitly below.
Recent progress in dense optical flow has been driven by increasingly complex architectures and multi-step refinement for test-time scaling. While these approaches achieve strong benchmark performance, they also require substantial computation during inference. This raises a fundamental question: Is scaling test-time computation the only way to improve dense optical flow accuracy? We argue that it is not. Instead, powerful visual semantic and geometric priors encoded in modern foundation models can reduce, if not overcome, the need for computationally expensive iterative refinement at test-time. In this paper, we present a framework that estimates dense optical flow in a single forward pass, leveraging pretrained foundation representations, while avoiding iterative refinement and additional inference-time computation, thus offering an alternative to test-time scaling. Our method extracts visual semantic features from a frozen DINO-v2 backbone and combines them with geometric cues from a monocular depth foundation model. We fuse these complementary priors into a unified representation and apply a global matching formulation to estimate dense correspondences without recurrent updates or test-time optimization. Despite avoiding iterative refinement, our approach achieves strong cross-dataset generalization across challenging benchmarks. On Sintel Final, we obtain 2.81 EPE without refinement, significantly improving over state-of-the-art (SOTA) SEA-RAFT under comparable training conditions and outperforming RAFT, GMFlow (without refinement), and recent FlowSeek in the same setting. These results suggest that strong foundation priors can substitute for test-time scaling, offering a computationally efficient alternative to refinement-heavy pipelines.
Praroop Chanda, Suryansh Kumar
Visual and Spatial AI Lab1, VCCM Section College of PVFA2, Department of ECEN3, Department of CSCE4, Texas A&M University, College Station, Texas, USA
Optical flow methods typically rely on task-specific inductive biases, such as correlation volumes, feature warping, and iterative refinement, among others, to reach high accuracy. While effective, such biases constrain the model to predefined heuristics, which can limit its expressivity and lead to more complex pipelines and additional computational cost. We present FreeFlow, a hierarchical transformer built without any flow-specific components, using instead a single feed-forward encoder--decoder. FreeFlow combines three attention variants: window attention for local processing, shifted-window attention for cross-window information exchange, and a global attention operating at a reduced resolution. The resulting architecture scales naturally with model capacity, enabling a consistent accuracy gain from small to large variants. Despite the absence of standard inductive biases, FreeFlow achieves state-of-the-art results on major benchmarks, including Sintel (0.68/1.48 EPE on Clean/Final), KITTI-2015 (3.23 Fl-all), and Spring (3.192 1px), while remaining memory efficient at 1080p inference.
Vladislav Bargatin, Alexander Yakovenko, Khaled Abud +1
AI Center, Lomonosov MSU, Moscow, Russia · Lomonosov Moscow State University, Moscow, Russia · MSU Institute for Artificial Intelligence, Moscow, Russia
We present FlowIt, a novel architecture for optical flow estimation that combines global matching with confidence and occlusion-guided refinement. At its core, FlowIt leverages a hierarchical transformer architecture that captures extensive global context, enabling the model to effectively model long-range correspondences. To overcome the limitations of localized matching, we formulate the flow initialization as an optimal transport problem. This formulation yields a highly robust initial flow field, alongside explicitly derived occlusion and confidence maps. These cues are then seamlessly integrated into a guided refinement stage, where the network actively propagates reliable motion estimates from high-confidence regions into ambiguous, low-confidence areas. Extensive experiments across the Sintel, KITTI, Spring, and LayeredFlow datasets validate the effectiveness of our approach. FlowIt achieves state-of-the-art results on the competitive Sintel benchmark and establishes new state-of-the-art cross-dataset zero-shot generalization performance on Sintel, Spring, and LayeredFlow, while also delivering competitive performance on both the KITTI benchmark and KITTI zero-shot generalization settings.
Sadra Safadoust, Fabio Tosi, Matteo Poggi +1
Department of Computer Engineering and KUIS AI Center, Koç University, Istanbul, Turkey · Department of Computer Science and Engineering (DISI), University of Bologna, Italy