Mobility Helps Learning: Unsupervised Model Adaptation for Object Recognition via Movement

Yanan Ma; Yihang Tao; Zhengru Fang; Senkang Hu; Xianhao Chen; Yuguang Fang

cs.CVOct 11, 2026

Mobility Helps Learning: Unsupervised Model Adaptation for Object Recognition via Movement

Authors: Yanan Ma, Yihang Tao, Zhengru Fang, Senkang Hu, Xianhao Chen, Yuguang Fang

Organizations: Department of Computer Science, City University of Hong Kong · Department of Electrical and Computer Engineering, The University of Hong Kong

Abstract

Movable agents, such as autonomous vehicles and robots, should be able to autonomously adapt their pre-trained object recognition models to new environments without human annotations. While existing test-time adaptation (TTA) methods address label-free model adaptation, they typically treat test data as independent snapshots, without constructively exploiting a rich, unique source of supervision: the significant variance in prediction quality as an agent observes the same object from different distances and viewpoints. To capitalize on this, we propose MoCaFe, a model-agnostic and hyperparameter-insensitive paradigm that leverages motion-induced prediction variance for unsupervised adaptation. Specifically, we first develop mobility-calibrated filtering of pseudo-labels, which constructs reliable pseudo-labels by fusing cross-view predictions using inverse-variance weights. The filtered pseudo-labels then guide the selection of ``hard samples'' for effective learning. To further suppress pseudo-label corruption, we introduce p-MoCaFe, an optional extension that uses a public dataset to assign weighting factors to data samples, yielding an estimator of the clean loss with bounded bias. Extensive experiments on autonomous driving (nuScenes, KITTI) and embodied-agent datasets show that MoCaFe significantly outperforms state-of-the-art TTA schemes, achieving up to approximately 20% accuracy gain.

Figures & tables

Panel of Figure 1 — Figure 1: (a) The agent is more likely to produce the correct class, i.e., cat, when moving closer to the object. Intuitively, the high-confidence predictions (on the right) can serve as pseudo-labels to guide model adaptation over samples with lower confidence (on the left). (b) Average entropy and classification accuracy versus the distance for two classes of objects, i.e., barriers and motorcycles, in the nuScenes dataset with a pre-trained ResNet-50. Accuracy varies significantly as agents move .

Figure 2: Overview of the MoCaFe and p-MoCaFe frameworks detailed in Section 4 . Multi-view predictions are aggregated via reliability weighting to construct a robust "Filtered Teacher" pseudo-label. Discrepancy scores against this teacher are then used to select the most informative student samples for training. In the p-MoCaFe extension, an additional public dataset is used to calibrate debiasing weights, yielding an unbiased empirical risk for robust model adaptation.

	ResNet-50			Swin-T
Pre-trained	75%	70%	65%	75%	70%	65%
w/o MoCaFe	76.18 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	72.05 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	62.17 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.37)	76.69 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	72.55 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	40.23 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.43)
AdaCon	77.88 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.19)	73.63 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.15)	66.51 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.18)	76.85 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.33)	72.83 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.47)	62.28 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.61)
PLUE	77.94 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.12)	73.82 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.21)	67.28 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.29)	77.43 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	73.55 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.30)	59.91 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.35)
RoTTA	78.67 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.21)	75.13 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.33)	68.20 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.31)	—–	—–	—–
C-SFDA	78.71 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.22)	75.37 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	68.34 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.34)	77.75 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.18)	72.20 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.19)	66.43 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)

Table 1: Performance in the Boston Seaport scenario on the nuScenes dataset under varying pre-trained model accuracy (i.e., 75%, 70%, and 65%). The value in the bracket represents the standard deviation. Boldface and Underline highlight the highest and second-highest accuracy, except for the FSL benchmark.

\color[rgb]{0,1,0}\uparrow

and

\color[rgb]{1,0,0}\downarrow

indicate performance improvement/degradation.

	ResNet-50
Class	Car	Pedestrian	Traffic-Cone	Trailer	Bus	Motorcycle	Barrier	Con-Vehicle	Bicycle	Truck
Pre-trained	82.42	77.35	70.46	77.22	62.86	61.11	53.09	49.64	48.15	23.28
w/o MoCaFe	87.41 $\color[rgb]{0,1,0}\uparrow$	81.05 $\color[rgb]{0,1,0}\uparrow$	76.95 $\color[rgb]{0,1,0}\uparrow$	75.32 $\color[rgb]{1,0,0}\downarrow$	61.90 $\color[rgb]{1,0,0}\downarrow$	47.22 $\color[rgb]{1,0,0}\downarrow$	52.31 $\color[rgb]{1,0,0}\downarrow$	49.68 $\color[rgb]{0,1,0}\uparrow$	46.91 $\color[rgb]{1,0,0}\downarrow$	18.84 $\color[rgb]{1,0,0}\downarrow$
AdaCon	88.38 $\color[rgb]{0,1,0}\uparrow$	85.29 $\color[rgb]{0,1,0}\uparrow$	77.42 $\color[rgb]{0,1,0}\uparrow$	81.00 $\color[rgb]{0,1,0}\uparrow$	68.10 $\color[rgb]{0,1,0}\uparrow$	50.05 $\color[rgb]{1,0,0}\downarrow$	53.09 $\color[rgb]{0.5,0.5,0.5}\rightarrow$	40.53 $\color[rgb]{1,0,0}\downarrow$	37.50 $\color[rgb]{1,0,0}\downarrow$	14.64 $\color[rgb]{1,0,0}\downarrow$
PLUE	88.79 $\color[rgb]{0,1,0}\uparrow$	84.98 $\color[rgb]{0,1,0}\uparrow$	75.55 $\color[rgb]{0,1,0}\uparrow$	78.67 $\color[rgb]{0,1,0}\uparrow$	64.87 $\color[rgb]{0,1,0}\uparrow$	55.89 $\color[rgb]{1,0,0}\downarrow$	54.24 $\color[rgb]{0,1,0}\uparrow$	46.15 $\color[rgb]{1,0,0}\downarrow$	39.71 $\color[rgb]{1,0,0}\downarrow$	15.51 $\color[rgb]{1,0,0}\downarrow$
RoTTA	87.41 $\color[rgb]{0,1,0}\uparrow$	83.78 $\color[rgb]{0,1,0}\uparrow$	73.12 $\color[rgb]{0,1,0}\uparrow$	74.74 $\color[rgb]{1,0,0}\downarrow$	65.39 $\color[rgb]{0,1,0}\uparrow$	52.53 $\color[rgb]{1,0,0}\downarrow$	58.33 $\color[rgb]{0,1,0}\uparrow$	40.53 $\color[rgb]{1,0,0}\downarrow$	37.62 $\color[rgb]{1,0,0}\downarrow$	12.87 $\color[rgb]{1,0,0}\downarrow$

Table 2: Per-class test accuracy in the Boston Seaport scenario on the nuScenes dataset. Pre-trained accuracy for the Boston scenario is 70%.

\color[rgb]{0.5,0.5,0.5}\rightarrow

represents that the performance remains the same with the pre-trained model.

	ResNet-50			Swin-T
Pre-trained	75%	70%	65%	75%	70%	65%
w/o MoCaFe	75.53 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	71.41 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	61.48 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.39)	76.10 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	71.96 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.27)	39.51 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.39)
AdaCon	76.45 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.14)	72.72 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.18)	65.46 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.19)	76.23 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	72.16 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.22)	60.39 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.54)
PLUE	77.16 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.16)	73.04 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	66.80 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	76.88 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	73.00 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.32)	57.58 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.34)
RoTTA	77.35 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	73.85 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.23)	67.09 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.27)	—	—	—
C-SFDA	78.03 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.21)	73.99 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	67.60 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.31)	77.11 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.20)	71.59 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.23)	65.42 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.27)

Table 3: Performance in the Boston Seaport scenario on the nuScenes dataset with sequences constructed by an off-the-shelf object tracking algorithm.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Pre-train Acc.	Pseudo-labeling Strategy	PL Correctness (%)	Test Acc. (%)
75.0%	Max-confidence	84.21	80.24
75.0%	Filtered (proposed)	84.95	80.81
42.4%	Max-confidence	42.13	41.10
42.4%	Filtered (proposed)	44.80	44.07
37.4%	Max-confidence	34.85	33.12
37.4%	Filtered (proposed)	37.73	37.04

Appendix

Table 4: Robustness of pseudo-label construction. We investigate the pseudo-label construction module in MoCaFe with ResNet-50 under the Boston Seaport scenario. “PL correctness” measures the correctness percentage of constructed pseudo-labels.

Pre-train Acc.	Pseudo-labeling Strategy	PL Correctness (%)	Test Acc. (%)
75.0%	Max-confidence	86.54	88.81
75.0%	Filtered (proposed)	88.72	89.60
57.0%	Max-confidence	64.68	68.23
57.0%	Filtered (proposed)	68.35	72.80
51.0%	Max-confidence	58.80	61.56
51.0%	Filtered (proposed)	63.76	67.57

Appendix

Table 5: Robustness of pseudo-label construction. We investigate the pseudo-label construction module in MoCaFe with Swin-T under the Boston Seaport scenario.

Data Sampling	Data Weighting	ResNet-50			Swin-T
Data Sampling	Data Weighting	75%	70%	65%	75%	70%	65%
✗	✔	77.24 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.20)	73.66 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	65.48 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.29)	77.80 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	74.82 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	43.98 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.33)
✔	✗	80.81 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.23)	77.25 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	69.32 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	89.60 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.11)	89.34 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.15)	89.27 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.16)
✔	✔	81.23 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.13)	77.93 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.16)	69.74 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	90.19 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.17)	90.07 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.18)	90.02 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.12)

Appendix

Table 6: Ablation study on key components of p-MoCaFe in the Boston Seaport scenario on the nuScenes dataset with best in bold .

Table 9

Table 10

Panel of Figure 3 — Figure 3: The test accuracy versus epoch with the object detection model Faster R-CNN.

	ResNet-50			Swin-T
Num. Samples	$m=32000$	$m=24000$	$m=16000$	$m=32000$	$m=24000$	$m=16000$
w/o MoCaFe	76.18 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	75.36 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.15)	75.31 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.18)	76.69 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	76.06 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	75.87 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)
AdaCon	77.88 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.19)	76.93 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.23)	76.27 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	76.85 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.33)	76.24 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.35)	75.73 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.37)
PLUE	77.94 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.12)	76.48 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.23)	75.84 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	77.43 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	76.95 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	76.51 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)
RoTTA	78.67 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.21)	77.93 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.30)	77.22 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.29)	—–	—–	—–
C-SFDA	78.71 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.22)	77.54 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.27)	76.63 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.30)	77.75 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.18)	77.36 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.20)	76.97 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.22)

Appendix

Table 11: Performance in the Boston Seaport scenario on the nuScenes dataset under different numbers of selected training samples

m

. The pre-trained model accuracy for the Boston scenario is 75%. Boldface and Underline highlight the highest and second-highest accuracy, except for the fully-supervised fine-tuning benchmark.

	ResNet-50			Swin-T
Pre-train	75%	70%	65%	75%	70%	65%
w/o MoCaFe	76.20 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	72.14 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	60.46 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.43)	75.98 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.31)	72.24 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	41.42 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.38)
AdaCon	77.03 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.30)	73.38 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	60.74 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.32)	76.32 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	73.51 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.30)	59.78 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.43)
PLUE	77.56 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	73.05 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.31)	67.32 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.35)	76.88 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	73.40 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.27)	65.32 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.34)
RoTTA	78.16 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	74.51 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	67.75 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.39)	—–	—–	—–
C-SFDA	78.00 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.20)	74.98 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.32)	67.79 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	77.15 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.23)	72.87 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	67.64 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.29)

Appendix

Table 12: Performance on the KITTI dataset under varied pre-trained model accuracy (i.e., 75%, 70%, and 65%). Boldface and Underline highlight the highest and second-highest accuracy, except for the fully-supervised fine-tuning benchmark.

	ResNet-50			Swin-T
Num. Samples	$m=32000$	$m=24000$	$m=16000$	$m=32000$	$m=24000$	$m=16000$
w/o MoCaFe	76.20 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	75.82 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.34)	75.37 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.31)	75.98 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.31)	74.71 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.23)	74.24 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.24)
AdaCon	77.03 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.30)	76.75 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.21)	75.89 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	76.32 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	75.94 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	75.23 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.34)
PLUE	77.56 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	77.11 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	76.74 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.20)	76.88 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	76.51 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.22)	75.86 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)
RoTTA	78.16 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	77.72 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	77.30 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.23)	—–	—–	—–
C-SFDA	78.00 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.20)	77.53 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.22)	77.13 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.27)	77.15 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.23)	76.73 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	76.20 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.29)

Appendix

Table 13: Performance on the KITTI dataset under different numbers of selected training samples

m

. The pre-trained model accuracy on the KITTI dataset is 75%.

	ResNet-50			Swin-T
Pre-train	75%	70%	65%	75%	70%	65%
w/o MoCaFe	76.03 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	71.06 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.29)	60.07 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.34)	76.03 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)	71.32 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.31)	38.47 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.35)
AdaCon	76.85 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	72.63 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	65.63 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.29)	76.81 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.22)	72.04 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.24)	58.48 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.27)
PLUE	77.01 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.25)	72.88 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.27)	65.75 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.33)	76.83 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.19)	72.79 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.21)	60.27 $\color[rgb]{1,0,0}\downarrow$ ( $\pm$ 0.30)
RoTTA	77.76 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.18)	74.00 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.20)	67.36 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.23)	—-	—-	—-
C-SFDA	77.93 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.22)	73.66 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.22)	68.53 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.31)	77.12 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.17)	73.02 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.26)	66.88 $\color[rgb]{0,1,0}\uparrow$ ( $\pm$ 0.28)

Appendix

Table 14: Performance on the embodied agent scenario under varied pre-trained model accuracy for this scenario (i.e., 75%, 70%, and 65%). The public dataset is still the Singapore Onenorth scenario on the nuScenes dataset.

Test-Time Adaptation

Test-time adaptation updates or recalibrates pretrained models at inference to handle distribution shifts, often without target labels.

See all 322 papers

Learning from Distributed Eyes: Leveraging Collaborative Perception for Automated Model Adaptation
Sep 16, 2026Yanan Ma, Yihang Tao, Zhengru Fang +4Unsupervised Domain Adaptation Autonomous Driving Perception
Hong Kong JC Lab of Smart City and the Department of Computer Science, City University of Hong Kong, Hong Kong, China · School of Data Science, Lingnan University, Tuen Mun, Hong Kong, China · Department of Electrical and Electronic Engineering, University of Hong Kong, Hong Kong, China
Dual-level Adaptation for Multi-Object Tracking: Building Test-Time Calibration from Experience and Intuition
Mar 23, 2026Wen Guo, Pengfei Zhao, Zongmeng Wang +2Multi-Object Tracking Test-Time Adaptation
Shandong Technology and Business University · Inner Mongolia University · University of Science and Technology Beijing +1
Class-Incremental Motion Forecasting
Mar 10, 2026Nicolas Schischka, Nikhil Gosala, B Ravi Kiran +2Class-Incremental Learning Autonomous Driving Perception
Department of Computer Science, University of Freiburg, Germany. · Qualcomm SARL France. · Automated Driving, Qualcomm Technologies, Inc.

Sep 16, 2026J/K move · Enter open · S save

Learning from Distributed Eyes: Leveraging Collaborative Perception for Automated Model Adaptation

Yanan Ma, Yihang Tao, Zhengru Fang, Zihan Fang+3

Hong Kong JC Lab of Smart City and the Department of Computer Science, City University of Hong Kong, Hong Kong, China · School of Data Science, Lingnan University, Tuen Mun, Hong Kong, China · Department of Electrical and Electronic Engineering, University of Hong Kong, Hong Kong, China

Unsupervised Domain Adaptation Autonomous Driving Perception 3D Object Detection Pseudo-Labeling

Open paper AskPDF arXiv

In autonomous driving, perception models often struggle to generalize to new environments due to domain shifts. While unsupervised model adaptation offers a feasible solution without labor-intensive manual labeling, existing methods that rely solely on the ego-vehicle's data often lead to inferior pseudo-labeling performance. To address this critical issue, we propose LDE, Learning from Distributed ``Eyes", a novel framework that transforms collaborative perception (CP) into a source of high-quality supervision for model adaptation. This pseudo-labeling approach is hyperparameter-insensitive and relatively reliable, assuming CP often outperforms single-agent's perception. However, naively implementing this approach encounters (1) the communication bottleneck of sharing rich features under time and bandwidth constraints, (2) the view discrepancy between the CP view and the learner's Field of View (FoV), and (3) the unreliability even in CP-generated labels. To address these issues, we design an adaptation-oriented feature sharing mechanism that selectively transmits the most critical information for adaptation, an FoV filtering method that meticulously eliminates mismatched labels, and a curriculum learning strategy to progressively exploit pseudo labels. Extensive experiments on 3D object detection tasks demonstrate that LDE consistently outperforms both the pre-trained models and state-of-the-art unsupervised adaptation methods.

Abstract

Test-Time Adaptation

Explore similar work

Learning from Distributed Eyes: Leveraging Collaborative Perception for Automated Model Adaptation

Dual-level Adaptation for Multi-Object Tracking: Building Test-Time Calibration from Experience and Intuition

Class-Incremental Motion Forecasting