Stable and reliable 4D spatial understanding is fundamental for autonomous driving systems. While feedforward reconstruction networks can estimate camera motion and 3D structure in one pass, pose estimation over long videos remains challenged by computational cost, long-context ambiguity, and temporal instability. To address these challenges, we propose Feedforward Visual Odometry (FFVO), a pose-specialized adaptation of joint reconstruction architectures for efficient and temporally stable camera-pose estimation. FFVO uses (i) a compact camera-token representation for computationally efficient temporal aggregation, (ii) a hierarchical local-to-global temporal decoder that mitigates geometric ambiguity by separating short-range motion aggregation from sequence-level integration, and (iii) intermediate trajectory supervision that promotes temporal stability. Extensive evaluation on the Waymo Open Dataset (WOD), KITTI, and a large-scale proprietary benchmark demonstrates that our method performs favorably against existing feedforward approaches, and greatly reduces jitter and drift. These results support FFVO as an effective feedforward camera-pose decoder in long-horizon visual odometry settings.
Figures & tables
Figure 2: Feedforward reconstruction. As shown in (b), the state-of-the-art feedforward model π3 [ 40 ] is able to produce locally accurate geometry. However, it exhibits pose jitter that misaligns geometry across frames, leading to duplicated static structures (e.g., road signs and barriers) when given long-horizon videos as input, as shown in (c). This observation motivates us to develop a better pose regression branch to improve long term temporal consistency. As shown in (d), our reconstruction is more temporally coherent. Camera poses: yellow denotes the current frame, red denotes other frames. We remove dynamic agents and sky for better visualization.
Figure 3: Overview. A frozen π3 [ 40 ] backbone produces image tokens and geometry, while the trainable camera-token and local-to-global pose branches predict long-horizon odometry using relative-pose and intermediate local-trajectory supervision.
Figure 4: Camera tokenization. For each frame t , we partition the frozen image tokens Xt={xt(j)}j=1N into M coarse regions Ri and average-pool xt(j) within each region to initialize camera tokens Ct={ct(i)}i=1M , which serve as compact pose queries for temporal cross-attention.
Figure 5: Local-to-global temporal pose decoder with its overall architecture and key attention modules. (a) Overall architecture with two local-window stages and one global integration stage. (b) Frame self-attention within a single frame. (c) Temporal cross-attention where camera tokens query temporal context.
Sequences from the PDD test set with fast, long trajectories.
Figure 7: Qualitative comparison on KITTI Odometry Seq. 02. The blue and red boxes highlight local trajectory geometry: π3 -Long shows a local self-collision in the blue box, and both π3 -Long and DA3-Streaming produce overly-close/merged curves in the red box, while Ours-Long better preserves the separation and overall shape compared to GT.
Figure 8: Failure modes. Our model shows similar patterns of temporal instability and inaccuracy as other feedforward methods under degraded conditions: (top) high-speed motion exceeding 55 mph, (center) extreme weather, and (bottom) low-light environments.
Methods
Avg.
00
01
02
03
04
05
06
07
08
09
10
Classical
ORB-SLAM2 w/o LC [ 24 ] †
69.727
40.65
502.20
47.82
0.94
1.30
29.95
40.82
16.04
43.09
38.77
5.42
ORB-SLAM2 w/ LC [ 24 ] †
54.816
6.03
508.34
14.76
1.02
1.57
4.04
11.16
2.19
38.85
8.39
6.63
LDSO [ 10 ] †
22.425
9.32
11.68
31.98
2.85
1.22
5.10
13.55
2.96
129.02
21.64
17.36
Learning Based
DROID-VO [ 32 ] †
54.188
98.43
84.20
108.80
2.58
0.93
59.27
64.40
24.20
64.55
71.80
16.91
DROID-SLAM [ 32 ] †
100.278
92.10
344.60
107.61
2.38
1.00
118.50
62.47
21.78
161.60
72.32
118.70
DPVO [ 33 ] †
53.609
113.21
12.69
123.40
2.09
0.68
58.96
54.78
19.26
115.90
75.10
13.63
Table 3: KITTI Odometry eval results . Top-2 highlighting excludes classical methods. † : from [ 6 ] . ‡ : per-sequence values are reported where available.
Figure 9: Trajectory visualization on the KITTI Odometry Benchmark. Predictions and ground truth for sequences 00–10; all methods use 120-frame alignment chunks [ 6 ] .
Method
ATE (m) ↓
RPE t (m) ↓
RPE r (deg) ↓
π3 -finetuned
15.989
8.452
0.596
+ camera tokens (cross-attn., M=14 )
11.776 (+26.35%)
5.509 (+34.82%)
0.379 (+36.41%)
+ auxiliary supervision
11.195 (+29.98%)
5.354 (+36.65%)
0.373 (+37.42%)
+ local-window temporal attention
7.070 (+55.78%)
1.090 (+87.10%)
0.289 (+51.51%)
+ increase camera tokens to M=24
7.060 (+55.84%)
1.070 (+87.34%)
0.281 (+52.85%)
Table 4: Ablation study on PDD. All our proposed components contribute meaningfully to the final performance. Green number indicates relative improvement to π3 -finetuned.
Robotics and Autonomous Systems Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou 511453, China · Cheng Kar-Shun Robotics Institute, The Hong Kong University of Science and Technology, Hong Kong SAR, China