Low-visibility scenarios, such as low-light conditions, pose significant challenges to human pose estimation due to the scarcity of annotated low-light datasets and the loss of visual information under poor illumination. Recent domain adaptation techniques attempt to utilize well-lit labels by augmenting well-lit images to mimic low-light conditions. But handcrafted augmentations oversimplify noise patterns, while learning-based methods often fail to preserve high-frequency low-light characteristics, producing unrealistic images that lead pose models to generalize poorly to real low-light scenes. Moreover, recent pose estimators rely on image cues through image-to-keypoint cross-attention, but these cues become unreliable under low-light conditions. To address these issues, we propose Unsupervised Domain Adaptation for Pose Estimation (UDAPose), a novel framework that synthesizes low-light images and dynamically fuses visual cues with pose priors for improved pose estimation. Specifically, our synthesis method incorporates a Direct-Current-based High-Pass Filter (DHF) and a Low-light Characteristics Injection Module (LCIM) to inject high-frequency details from input low-light images, overcoming rigidity or the detail loss in existing approaches. Furthermore, we introduce a Dynamic Control of Attention (DCA) module that adaptively balances image cues with learned pose priors in the Transformer architecture. Experiments show that UDAPose outperforms state-of-the-art methods, with notable AP gains of 10.1 (56.4%) on the ExLPose-test hard set (LL-H) and 7.4 (31.4%) in cross-dataset validation on EHPT-XC. Code: https://github.com/Vision-and-Multimodal-Intelligence-Lab/UDAPose
Figures & tables
Figure 2: Limitations of learning-based low-light augmentation. The first two columns show well-lit and paired low-light images from ExLPose [ 30 ] . The third and fourth columns present results from CycleGAN [ 80 ] and StyleID [ 9 ] . The last column shows our result. Low-light images are scaled to an average channel intensity of 0.4 for visualization only.
Figure 3: Overview of the UDAPose framework. During augmentation, the LCIM uses extracted low-light features from unpaired low-light images ( ILL ) to synthesize low-light counterparts ( I^LL ) of well-lit images ( IWL ). These synthetic images retain the original pose annotations from IWL while accurately reflecting low-light characteristics of ILL . The pose model is trained using these augmented images and their inherited annotations. Our DCA adaptively balances image cues and pose priors under low-light conditions. During inference, the trained model is directly applied to real low-light images. Note that ILL , ILL′ , and I^LL are scaled for visualization only.
Figure 4: Ratio of Frobenius Norm of Qimage over Qpose (i.e. ∥Qimage∥2/∥Qpose∥2 ) on different keypoints and pose estimation results before/after applying DCA module. Note that images are scaled for visualization only.
Methods
AP ↑ @.50:.95
WL
LL-A
LL-N
LL-H
LL-E
RFormer [ 3 ]
60.0
4.5
15.2
0.3
0.8
DarkIR [ 13 ]
60.2
6.1
17.4
1.3
1.2
LightenDiff [ 19 ]
60.1
5.6
13.9
0.7
0.8
QuadPrior [ 65 ]
60.2
8.9
19.3
4.6
0.3
CycleGAN [ 80 ]
61.3
19.6
33.7
17.9
3.3
Table 1: Evaluation mAP on ExLPose-test comparing image enhancement and domain adaptation methods. All methods are trained only on augmented images and well-lit annotations, without using low-light ground truth or paired data. The best is bold . The second best is underlined .
Methods
AR ↑ @.50:.95
WL
LL-A
LL-N
LL-H
LL-E
RFormer [ 3 ]
71.5
9.7
25.4
4.1
0.6
DarkIR [ 13 ]
72.0
11.1
29.3
7.3
0.8
LightenDiff [ 19 ]
71.8
10.2
22.9
4.9
0.5
QuadPrior [ 65 ]
71.9
15.5
30.9
11.7
1.0
CycleGAN [ 80 ]
72.1
28.7
45.6
27.4
8.9
Table 2: Evaluation mAR on ExLPose-test comparing image enhancement and domain adaptation methods, following Tab. 1 . The best is bold . The second best is underlined .
Figure 5: Qualitative comparisons of our method with existing baselines, including image enhancement [ 65 ] , domain adaptation [ 1 ] , and image translation method [ 26 ] . From top to bottom, we show samples from low-light normal, hard, and extreme sets from ExLPose-test [ 30 ] . Our approach consistently outperforms the existing methods in human pose estimation across all scenarios. Images are scaled up for visualization purpose only.
Methods
AP ↑ @.50:.95
AR ↑ @.50:.95
Avg.
A7 M3
RIC OH3
Avg.
A7 M3
RIC OH3
RFormer [ 3 ]
27.5
29.4
25.7
43.7
47.2
40.3
DarkIR [ 13 ]
28.9
30.3
27.6
47.0
48.7
45.4
LightenDiff [ 19 ]
25.3
29.1
21.5
40.6
45.4
35.9
QuadPrior [ 65 ]
29.3
30.6
28.0
48.8
50.4
47.2
CycleGAN [ 80 ]
45.1
47.5
42.7
60.7
63.7
57.8
Table 3: Evaluation on ExLPose-OCN, following identical setup as in Tab. 1 . The best is bold . The second best is underlined .
Methods
AP ↑
AR ↑
@.50:.95
@.50
@.75
@.50:.95
@.50
@.75
RFormer [ 3 ]
8.8
20.3
7.9
21.3
39.2
18.9
DarkIR [ 13 ]
12.5
25.2
11.3
28.9
48.5
27.1
LightenDiff [ 19 ]
9.7
17.7
8.3
22.0
38.4
20.3
QuadPrior [ 65 ]
12.9
24.5
11.5
28.1
47.7
27.5
CycleGAN [ 80 ]
20.7
36.9
19.2
45.0
69.6
45.5
Table 4: Cross-dataset validation on EHPT-XC [ 8 ] , using the model weights as in Tab. 1 . Best is bold , second best underlined .
AP ↑ @0.5:0.95
WL
LL-N
LL-H
LL-E
A7 M3
RIC OH3
EHPT -XC
Well-lit
60.1
3.4
0.4
0.2
11.3
13.4
0.3
HM
60.0
21.3
1.3
0.5
15.1
11.4
1.5
Baseline SD
60.0
23.7
7.2
0.0
30.1
25.4
3.2
+ AIN
60.1
25.2
8.8
2.4
33.0
27.8
6.1
+ LCIM
60.1
31.5
20.7
7.8
43.1
39.8
19.5
Table 5: Ablation study of our proposed modules on ExLPose-test, ExLPose-OCN, and EHPT-XC. Well-lit: pose model trained with well-lit images only. HM: pose model adapted with synthetic low-light images using histogram matching. AIN is a normalization step (see supplementary). The best is bold .
AP ↑ @0.5:0.95
WL
LL-N
LL-H
LL-E
A7 M3
RIC OH3
EHPT -XC
Ours (ExLPose)
67.3
38.7
28.0
11.7
55.0
47.9
31.0
Ours (CrowdPose)
71.7
47.9
37.2
19.4
70.5
69.7
45.7
Table 6: Comparison results of using ExLPose well-lit and CrowdPose to construct synthetic low-light training data.
Figure 6: The architecture of our DCA module.
AP ↑ @0.5:0.95
WL
LL-N
LL-H
LL-E
A7 M3
RIC OH3
EHPT -XC
4,000
66.1
34.7
22.4
5.4
50.0
45.1
25.4
8,000
66.3
35.4
24.8
7.8
53.2
46.8
27.9
12,000
66.9
36.6
26.2
10.4
54.7
47.2
29.1
16,000
66.8
37.5
27.3
11.3
55.0
47.8
30.5
20,000
67.3
38.7
28.0
11.7
55.0
47.9
31.0
Table 7: Performance vs. amount of synthetic training data.
Methods
PSNR ↑
SSIM ↑
LPIPS ↓
FID ↓
KL ↓
CycleGAN [ 80 ]
36.56
0.76
0.26
50.20
0.028
UNIT [ 39 ]
33.90
0.66
0.29
45.70
0.104
UNSB [ 26 ]
34.19
0.74
0.30
96.42
0.062
EnCo [ 2 ]
35.99
0.75
0.28
48.71
0.031
Ours
41.13
0.91
0.20
11.17
0.008
Table 8: Evaluation for human pose anatomical consistency of our method and learning-based baselines.
AP ↑ @0.5:0.95
WL
LL-N
LL-H
LL-E
A7 M3
RIC OH3
Main (ELLA)
62.1
29.4
13.6
1.6
35.0
27.2
Main (Ours)
61.5
32.3
23.2
8.3
37.2
35.0
Comp. (ELLA)
60.3
27.8
11.9
0.8
33.9
26.5
Comp. (Ours)
60.8
31.7
22.4
6.7
36.8
33.9
Student (ELLA)
60.8
35.6
18.6
5.0
39.1
36.2
Table 9: Full comparison results of ELLA [ 1 ] and our method. “Main” refers to “main teacher”. “Comp.” refers to “complementary teacher”. And “student” refers to “student” distillation model, which is the full model of ELLA. The best is bold .
AP ↑ @0.5:0.95
WL
LL-N
LL-H
LL-E
A7 M3
RIC OH3
Direct input
60.8
23.7
7.3
0.0
27.3
24.3
Z-score-based norm.
60.9
25.4
13.2
2.2
28.4
25.0
Fixed factor
58.1
29.0
20.8
6.4
33.2
31.0
ImageNet-based
61.3
31.7
22.4
7.4
36.5
33.8
Ours
61.5
32.3
23.2
8.3
37.2
35.0
Table 10: Evaluation of the AIN module on ExLPose-test and ExLPose-OCN. Direct input refers to feeding low-light images into SD without AIN. Experiments are conducted using the DEKR pose model [ 14 ] , with DHF and LCIM enabled for all normalization approaches. The best is bold .
AP ↑ @0.5:0.95
WL
LL-N
LL-H
LL-E
A7 M3
RIC OH3
EHPT -XC
z0
66.8
29.7
12.1
0.1
40.0
36.8
11.3
z0,z1
66.8
31.4
19.2
2.4
41.4
37.4
15.4
z0,z1,z2
67.2
35.3
23.2
5.8
43.8
39.9
26.2
z0,z1,z2,z3
67.4
37.7
26.5
7.7
47.9
43.7
29.7
z0,z1,z2,z3,z4
67.3
38.7
28.0
11.7
55.0
47.9
31.0
Table 11: Ablation study of LCIM on ExLPose-test, ExLPose-OCN, and EHPT-XC. z0 refers to baseline SD without any extra intermediate features. z1 to z4 represent low-to-high-frequency information fused in a coarse-to-fine integration strategy. Results are reported with AIN, DHF and DCA. The best is bold .
Figure 7: (a) Effect of λ . (b) Masking evaluation w/ and w/o DCA.
AP ↑ @0.5:0.95
WL
LL-N
LL-H
LL-E
A7 M3
RIC OH3
EHPT -XC
SE-Block [ 16 ]
62.4
36.7
26.3
9.5
50.3
46.5
26.7
CBAM [ 68 ]
62.5
37.0
26.2
9.8
51.1
46.2
27.0
Ours (DCA)
67.3
38.7
28.0
11.7
55.0
47.9
31.0
Table 12: Comparison of SE-Block [ 16 ] , CBAM [ 68 ] , and our DCA gating mechanism. The best is bold .
AP ↑ @0.5:0.95
WL
LL-N
LL-H
LL-E
A7 M3
RIC OH3
EHPT -XC
ControlNet [ 77 ]
66.3
31.7
16.4
2.7
47.6
43.7
22.4
IP-Adapter [ 76 ]
65.8
31.5
17.1
3.5
48.4
43.4
24.1
Ours
67.3
38.7
28.0
11.7
55.0
47.9
31.0
Table 13: Comparison of ControlNet [ 77 ] , IP-Adapter [ 76 ] , and our method. The best is bold .
Figure 8: Qualitative ablation of our DCA module. L. represents left, R. represents right.
Figure 9: Pose predictions of UDAPose compared against competing methods. The first two columns show results from enhancement-based methods; all other columns display results on the original low-light images. The low-light images are scaled for visualization only.
Unsupervised domain adaptation (UDA) enables robust transfer of knowledge from simulated to real environments while exploiting a subset of unlabeled target data to improve real-world performance. Existing UDA methods for Object pose estimation often rely on global feature matching, multi-stage larger frameworks, or image translation pipelines, which tend to overlook the pose-specific information embedded in feature representations. To bridge this limitation, we introduce CAPLR that targets the adaptation of pose-sensitive features in localized regions, ensuring that domain alignment preserves the geometric cues essential for accurate pose estimation. CAPLR achieves UDA with three key components: (1) Efficient Cross-Domain Pairing strategy leveraging intermediate features to identify pose similar image pairs across domains without supervision; (2) Contrastive Alignment to perform feature alignment at localised regions in both intermediate and task-specific representations; and (3) Consistency-Based Pseudo-Label Refinement to improve reliability by encouraging stable target predictions. Extensive experiments demonstrate that CAPLR achieves state-of-the-art performance across multiple well-known object pose estimation benchmarks featuring diverse and challenging scenarios.
Pre-trained diffusion models provide rich latent features across U-Net levels and are emerging as powerful vision backbones. While prior works such as Marigold and Lotus repurpose diffusion priors for dense geometric perception tasks such as depth and surface normal estimation, their potential for cross-domain human pose estimation remains largely unexplored. Through a systematic analysis of latent features from different upsampling levels of the Stable Diffusion U-Net, we identify the levels that deliver the strongest robustness and cross-domain generalization for pose estimation. Building on these findings, we propose SDPose, which (i) extracts U-Net features from the selected upsampling blocks, (ii) fuses them with a lightweight feature aggregation module to form a robust representation, and (iii) jointly optimizes keypoint heatmap supervision with an auxiliary latent reconstruction loss to regularize training and preserve the pre-trained generative prior. To evaluate cross-domain generalization and robustness, we construct COCO-OOD, a COCO-based benchmark with four subsets: three style-transferred splits to assess domain shift and one corruption split (noise, weather, digital artifacts, and blur) to test robustness. With data-efficient fine-tuning, SDPose achieves performance comparable to Sapiens on COCO, and establishes the best zero-shot performance on cross-domain datasets.
Shuang Liang, Jing He, Lejun Liao +4
Rama Alpaca Technology · The University of Hong Kong · HKUST (GZ) +3
We introduce Dark3R, a framework for structure from motion in the dark that operates directly on raw images with signal-to-noise ratios (SNRs) below −4 dB -- a regime where conventional feature- and learning-based methods break down. Our key insight is to adapt large-scale 3D foundation models to extreme low-light conditions through a teacher--student distillation process, enabling robust feature matching and camera pose estimation in low light. Dark3R requires no 3D supervision; it is trained solely on noisy--clean raw image pairs, which can be either captured directly or synthesized using a simple Poisson--Gaussian noise model applied to well-exposed raw images. To train and evaluate our approach, we introduce a new, exposure-bracketed dataset that includes ∼42,000 multi-view raw images with ground-truth 3D annotations, and we demonstrate that Dark3R achieves state-of-the-art structure from motion in the low-SNR regime. Further, we demonstrate state-of-the-art novel view synthesis in the dark using Dark3R's predicted poses and a coarse-to-fine radiance field optimization procedure.
Andrew Y Guo, Anagh Malik, SaiKiran Tedla +7
University of Toronto · Vector Institute · York University +3