CVPR 2026

UDAPose: Unsupervised Domain Adaptation for Low-Light Human Pose Estimation

Haopeng Chen¹ Yihao Ai² Kabeen Kim³ Robby T. Tan^2,4 Yixin Chen¹ Bo Wang¹

¹ University of Mississippi ² National University of Singapore ³ Duksung Women's University ⁴ ASUS Intelligent Cloud Services (AICS)

arXiv Paper (CVF) Code alphaXiv BibTeX

A real night test image of three people, brightened for display, with the closest detections drawn. — The well-lit model is trained on well-lit images only. Real night is ExLPose-OCN; Filtered is ExLPose-test.

The well-lit model is trained on well-lit images only. Real night is ExLPose-OCN; Filtered is ExLPose-test.

UDAPose learns pose estimation for the dark from well-lit labels alone. It synthesizes low-light training images from labeled well-lit ones, using unlabeled dark images as references, and adaptively fuses image cues with learned pose priors.

0low-light labels used

+10.1 AP(56.4%) on LL-H over the best baseline

11.7 vs 3.4AP on LL-E, more than 3x ELLA's augmentation in ED-Pose

+7.4 AP(31.4%) cross-dataset, EHPT-XC

Abstract

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: VMIL/UDAPose.

Samples

Prediction Ground truth

LL-N

LL-N test image, brightened for display.

LL-H

LL-H test image, brightened for display.

LL-E

LL-E test image, brightened for display.

A7M3

RICOH3

Comparison with other methods

Ground truth

LL-N

QuadPrior

LightenDiff

CycleGAN

ELLA

UDAPose

LL-N test image, brightened by QuadPrior.

LL-N test image, brightened by LightenDiff.

LL-N test image, brightened for display (CycleGAN column; skeletons need JavaScript).

LL-N test image, brightened for display (ELLA column; skeletons need JavaScript).

LL-N test image, brightened for display (UDAPose column; skeletons need JavaScript).

Swipe a row to see UDAPose. QuadPrior and LightenDiff brighten the test image before pose estimation, and their panels show the brightened result. CycleGAN darkens the well-lit training images. ELLA here is its low-light augmentation in ED-Pose; the full framework is under Questions. Every model was run on these images for this page.

Pose estimators go blind in the dark

A model trained on well-lit images scores 60.1 AP in good light. On the low-light subsets of ExLPose-test it scores 3.4, 0.4 and 0.2.

Table 5, row "Well-lit". AP @0.5:0.95 on ExLPose-test.

Labels for dark images barely exist, because people cannot see the joints either. Well-lit labels are plentiful. Can we reuse them?

Two routes, and what each one loses

Route A: brighten first

An LL-E image brightened by DarkIR, heavy noise, no skeleton on the person. Rendered by the authors. — DarkIR

An LL-E image brightened by QuadPrior, smooth and blurry, no skeleton on the person. Rendered by the authors. — QuadPrior

Enhance, then run a well-lit pose model. Recovering lost detail is ill-posed, and artifacts mislead the model.

Best enhancement AP on LL-H: 4.6 (QuadPrior, Table 1).

Route B: darken labeled images for training

Well-lit image darkened by CycleGAN. — CycleGAN

Well-lit image darkened by StyleID. Dark and smooth, with the scene visible. — StyleID

Keep the labels and train on synthetic darkness. Handcrafted noise oversimplifies: ELLA uses Gaussian noise. CycleGAN over-darkens and adds lighting artifacts; StyleID misses the noise.

Best translation AP on LL-H: 17.9 (CycleGAN, Table 1).

Left: an LL-E image rendered by the authors. Right: panels from the paper's Fig. 10, brightened for display.

The well-lit photo. — From the paper's Fig. 10. Each low-light panel is brightened for display by its own amount, and the UDAPose panels saturate more than the paired one, so judge the noise texture.

The paired low-light image, optically filtered, brightened for display, with strong color noise. — From the paper's Fig. 10. Each low-light panel is brightened for display by its own amount, and the UDAPose panels saturate more than the paired one, so judge the noise texture.

How UDAPose works

Well-lit photo of two people playing badminton in a sports hall, with their labeled skeletons. — 1. Normalized

Low-light street photo used as the style reference, brightened for display. — 1. Normalized

Insight 1

Low-light noise lives in the high frequencies

The latent path gives dark images with soft blotches instead of per-pixel speckle (step 2). So high frequencies get their own path: DHF extracts them from the reference, and LCIM puts them back. The payoff in Table 5: on top of Baseline SD, AIN and LCIM lift LL-H from 7.2 to 20.7 AP and LL-E from 0.0 to 7.8.

Clipping throws away the dark half

A high-pass image has a mean near zero, so about half its values are negative. The SD encoder expects values in [0, 1]. Clip directly and the dark half is gone. DHF shifts the signal first.

Adaptive Intensity Normalization (AIN) Scale the reference by one factor, toward a mean of 0.449 (Eq. 15).
High-pass FFT, mask, inverse FFT (Eq. 1).
Recenter Add back the difference of the means (Eq. 2).

I_DHF = I_HP + (mean(I_LL) − mean(I_HP))(2)

High-pass image clipped directly to [0, 1]. Most of it is dark, and the negative half is gone. — Computed with the released code (`util/fft.py`) on one LL-H reference. The panel is not brightened.

The recentered image. Gray, with both bright and dark fine detail. — Computed with the released code (`util/fft.py`) on one LL-H reference. The panel is not brightened.

LCIM puts them back, one scale at a time

Encoder features of I_DHF at four scales pass through a lightweight convolution, and each result is added at the end of a decoder block. Every added scale raises AP on every low-light column of Table 11.

LCIM is trained with the VAE frozen, to reconstruct low-light images with MSE plus a frequency loss weighted toward mid-to-high frequencies. Because its input is mostly noise, it captures characteristics that transfer to other images (Sec. 3).

L_D = L_MSE(I, Î) + λ L_freq(I, Î)(5)

Scales z0 to z4

Left: schematic. Right: Table 11. All rows include AIN, DHF and DCA, so "z0" is not the Table 5 "Baseline SD" row.

Table 11 and Table 10 in full

Table 11. LCIM scales, cumulative (AP @0.5:0.95, with AIN, DHF and DCA).
Method	WL	LL-N	LL-H	LL-E	A7M3	RICOH3	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 10. Normalizing the reference before the SD encoder (AP @0.5:0.95, DEKR setup).
Method	WL	LL-N	LL-H	LL-E	A7M3	RICOH3
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
UDAPose	61.5	32.3	23.2	8.3	37.2	35.0

Insight 2

A pose decoder cannot look away

DETR-style pose decoders add two signals with a fixed residual: Q_pose, the pose prior from self-attention, and Q_image, the image cues from cross-attention. In the dark the image cues are unreliable, and a fixed sum cannot back off.

Both knees barely visible

image, about 1.7

prior, 1

Well-lit images

image, about 1.68

prior, 1

Ratio of norms without DCA (Sec. 3). Barely visible knees stay comparable to clearly visible keypoints.

Person with a scooter. Without DCA the right knee sits out on empty background and the left knee sits down near the foot. — Before DCA: knees off the leg

Same person with DCA. Both knees sit on the leg. — Before DCA: knees off the leg

DCA lets each keypoint choose

Dynamic Control of Attention concatenates the two signals, runs a two-layer MLP, and applies a softmax. That gives two competing weights per keypoint. Their weighted sum replaces the fixed residual.

Redrawn from the paper's Fig. 6. Pose prior in teal, image cues in amber.

(w_pose, w_image) = softmax(MLP(Q_cat))(9)

Q = w_pose ⊙ Q_pose ⊕ w_image ⊙ Q_image(10)

The cost is two linear layers: 39 ms against 37.4 ms per image.

What it changes, joint by joint

Two cases from the paper's Fig. 8. The chart shows the norm ratio, not the DCA weights: lower means the image cues count for less against the prior. The ratio stays above 1, so DCA turns the image cues down without switching them off.

Before DCA: the left arm curls across the chest. — Values from the paper's Fig. 8, drawn on a schematic figure facing you.

After DCA: the left arm hangs straight. — Values from the paper's Fig. 8, drawn on a schematic figure facing you.

Case A. The arm across the dark shirt falls (L. elbow 1.72 to 1.41, L. wrist 1.69 to 1.24). The legs fall (knees about 1.85 to 1.55). The visible head holds (1.89 to 1.93).
Case B. The invisible knees and ankles fall: knees 1.83 to 1.13 and 1.86 to 1.17, ankles 1.88 to 1.31 and 1.87 to 1.35.
Not every drop maps to a hidden joint: in Case A the right wrist falls 1.74 to 1.49 with the hand visible.

All 14 joints for both cases

Fig. 8. Ratio of the norms of Q_image and Q_pose per joint.
Joint	Case A before	Case A after	Case B before	Case B after
L. shoulder	1.73	1.74	1.74	1.71
R. shoulder	1.81	1.78	1.71	1.74
L. elbow	1.72	1.41	1.82	1.73
R. elbow	1.77	1.47	1.81	1.72
L. wrist	1.69	1.24	1.85	1.74
R. wrist	1.74	1.49	1.84	1.93
L. hip	1.70	1.72	1.91	1.75
R. hip	1.80	1.81	1.93	1.73
L. knee	1.85	1.55	1.83	1.13
R. knee	1.84	1.57	1.86	1.17
L. ankle	1.76	1.51	1.88	1.31
R. ankle	1.81	1.52	1.87	1.35
Head	1.89	1.93	1.79	1.84
Neck	1.82	1.89	1.82	1.70

Does it help?

At the same decoder position, on the same synthetic data, DCA leads SE-Block and CBAM on every column of Table 12. It also helps in good light: AP on WL rises from 60.2 to 67.3 (Table 5).

Table 12. AP @0.5:0.95.

Table 12 in full

Table 12. Fusion gates at the same decoder position (AP @0.5:0.95).
Method	WL	LL-N	LL-H	LL-E	A7M3	RICOH3	EHPT-XC
SE-Block	62.4	36.7	26.3	9.5	50.3	46.5	26.7
CBAM	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

Line chart of mAP against the number of masked keypoints, 0 to 14, with and without DCA. With DCA stays higher until about 10 masked keypoints, where the lines meet. — The paper's Fig. 7b: random keypoints masked on A7M3. The gap narrows once most keypoints are masked, because with few visible joints the pose prior has little to build on.

Results

All methods use ED-Pose as the backbone. The UDAPose row also includes DCA, and Table 5 below separates the pieces.

ExLPose-OCN is real night; ExLPose-test is optically filtered. EHPT-XC is cross-dataset, with no retraining: 12 low-light scenes, 1200 images, selected from its merged train and test splits (Sec. 4). Hover, tap or focus a bar for the gap.

The same gap, in images

Input on an LL-N test image, from the paper's Fig. 5. — Input

QuadPrior on an LL-N test image, from the paper's Fig. 5. — Input

The darker, the larger the lead

In relative terms the lead grows with darkness: 1.15x, 1.56x, 3.44x. The absolute gain is 5.0, 10.1 and 8.3 AP, so it peaks on LL-H.

Table 1, best baseline per subset: CycleGAN on LL-N and LL-H, ELLA on LL-E.

What each piece adds

Table 5 adds the modules one at a time. Synthesis alone (the "+ DHF" row, no DCA) leads the best baseline by 7.4 AP on LL-H and by 0.8 on EHPT-XC. DCA adds the rest.

Table 5. AP @0.5:0.95.

Full results tables (Tables 1 to 5)

Table 1. AP on ExLPose-test (@.50:.95).
Method	WL	LL-A	LL-N	LL-H	LL-E
RFormer	60.0	4.5	15.2	0.3	0.8
DarkIR	60.2	6.1	17.4	1.3	1.2
LightenDiff	60.1	5.6	13.9	0.7	0.8
QuadPrior	60.2	8.9	19.3	4.6	0.3
CycleGAN	61.3	19.6	33.7	17.9	3.3
UNIT	54.1	7.4	16.1	3.6	0.9
UNSB	57.5	15.8	28.3	13.8	1.8
EnCo	60.0	17.2	31.9	16.2	2.9
UDA-HE	53.4	13.2	22.4	12.7	1.8
ELLA	61.5	18.6	32.3	17.2	3.4
UDAPose	67.3	27.0	38.7	28.0	11.7

Table 2. AR on ExLPose-test (@.50:.95).
Method	WL	LL-A	LL-N	LL-H	LL-E
RFormer	71.5	9.7	25.4	4.1	0.6
DarkIR	72.0	11.1	29.3	7.3	0.8
LightenDiff	71.8	10.2	22.9	4.9	0.5
QuadPrior	71.9	15.5	30.9	11.7	1.0
CycleGAN	72.1	28.7	45.6	27.4	8.9
UNIT	68.0	14.2	26.7	10.3	3.3
UNSB	69.4	23.8	39.7	22.5	5.7
EnCo	71.8	27.4	41.5	26.3	5.7
UDA-HE	67.2	21.4	32.7	20.9	5.8
ELLA	72.7	28.9	45.0	28.0	9.7
UDAPose	75.0	36.5	48.2	37.4	20.4

Table 3. ExLPose-OCN (@.50:.95).
Method	AP Avg.	AP A7M3	AP RICOH3	AR Avg.	AR A7M3	AR RICOH3
RFormer	27.5	29.4	25.7	43.7	47.2	40.3
DarkIR	28.9	30.3	27.6	47.0	48.7	45.4
LightenDiff	25.3	29.1	21.5	40.6	45.4	35.9
QuadPrior	29.3	30.6	28.0	48.8	50.4	47.2
CycleGAN	45.1	47.5	42.7	60.7	63.7	57.8
UNIT	35.1	40.7	29.6	55.2	61.9	48.6
UNSB	41.6	43.5	39.8	59.7	60.9	58.5
EnCo	43.2	45.4	41.1	58.4	61.5	55.4
UDA-HE	39.2	42.4	36.0	57.1	61.8	52.4
ELLA	46.0	48.5	43.5	62.3	65.5	59.1
UDAPose	51.4	55.0	47.9	65.1	68.1	62.2

Table 4. Cross-dataset validation on EHPT-XC.
Method	AP @.50:.95	AP @.50	AP @.75	AR @.50:.95	AR @.50	AR @.75
RFormer	8.8	20.3	7.9	21.3	39.2	18.9
DarkIR	12.5	25.2	11.3	28.9	48.5	27.1
LightenDiff	9.7	17.7	8.3	22.0	38.4	20.3
QuadPrior	12.9	24.5	11.5	28.1	47.7	27.5
CycleGAN	20.7	36.9	19.2	45.0	69.6	45.5
UNIT	11.2	24.5	8.7	35.5	64.2	32.5
UNSB	16.9	30.1	15.7	40.5	67.1	39.2
EnCo	18.3	33.4	17.1	42.3	68.1	42.8
UDA-HE	15.4	29.6	14.8	38.7	66.7	37.8
ELLA	23.6	38.3	22.7	48.2	73.2	48.6
UDAPose	31.0	51.1	30.3	51.3	76.4	53.8

Table 5. Ablation of the modules (AP @0.5:0.95).
Row	WL	LL-N	LL-H	LL-E	A7M3	RICOH3	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
+ DHF	60.2	35.3	25.3	9.4	48.9	45.6	24.4
+ DCA	67.3	38.7	28.0	11.7	55.0	47.9	31.0

Questions you might ask

Do the labels stay valid?

Predict pose heatmaps on a synthetic image and on its paired low-light image, then take the KL divergence. Lower means the joints stayed where the labels say. UDAPose scores 0.008 against 0.028 for the best baseline, with SSIM 0.91.

Table 8, KL. Heatmaps from DEKR.

Table 8 in full

Table 8. Anatomical consistency of the synthetic images.
Method	PSNR	SSIM	LPIPS	FID	KL
CycleGAN	36.56	0.76	0.26	50.20	0.028
UNIT	33.90	0.66	0.29	45.70	0.104
UNSB	34.19	0.74	0.30	96.42	0.062
EnCo	35.99	0.75	0.28	48.71	0.031
UDAPose	41.13	0.91	0.20	11.17	0.008

What if I have more well-lit data?

AP goes up. CrowdPose as the source, about 6x larger, takes A7M3 from 55.0 to 70.5 and RICOH3 from 47.9 to 69.7, close to its own well-lit AP of 71.7.

Table 6, AP. In Table 7, gains are larger below 12k synthetic images; the paper uses 20k.

Tables 6 and 7 in full

Table 6. ExLPose or CrowdPose as the well-lit source (AP @0.5:0.95).
Method	WL	LL-N	LL-H	LL-E	A7M3	RICOH3	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 7. AP vs amount of synthetic training data.
Method	WL	LL-N	LL-H	LL-E	A7M3	RICOH3	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

Why not ControlNet or IP-Adapter on paired data?

Both get pixel-level supervision from paired images, which UDAPose does not need. All rows have DCA on. UDAPose still leads on every column, by 10.9 AP on LL-H over the better of the two.

Table 13, AP on LL-H.

Table 13 in full

Table 13. Paired diffusion conditioning (AP @0.5:0.95).
Method	WL	LL-N	LL-H	LL-E	A7M3	RICOH3	EHPT-XC
ControlNet	66.3	31.7	16.4	2.7	47.6	43.7	22.4
IP-Adapter	65.8	31.5	17.1	3.5	48.4	43.4	24.1
UDAPose	67.3	38.7	28.0	11.7	55.0	47.9	31.0

How does it compare with full ELLA?

This test puts our synthetic data inside ELLA's own DEKR pipeline, without DCA. The student rises from 18.6 to 27.4 AP on LL-H. A supervised baseline with low-light labels, LSBN+LUPI, reaches 14.7.

Table 9, AP on LL-H.

Table 9 in full

Table 9. Inside the full ELLA framework (AP @0.5:0.95).
Method	WL	LL-N	LL-H	LL-E	A7M3	RICOH3
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
Student (Ours)	61.1	39.4	27.4	9.4	41.3	39.3
LSBN+LUPI	61.1	33.7	14.7	3.4	35.3	35.1

What are the limits?

It models darkness only. Fog, heavy rain or strong motion blur would need new synthesis modules.
Synthesis costs about 263 ms per image on an RTX 4090.
Extreme darkness is still hard: 11.7 AP on LL-E.

Citation

@InProceedings{chen2026udapose,
    author    = {Chen, Haopeng and Ai, Yihao and Kim, Kabeen and Tan, Robby T. and Chen, Yixin and Wang, Bo},
    title     = {UDAPose: Unsupervised Domain Adaptation for Low-Light Human Pose Estimation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {13781-13792}
}

arXiv Paper (CVF) Code alphaXiv

Acknowledgment

This work was supported in part by the Mississippi Impact Grant (MIG), Office for Research and Economic Development, University of Mississippi. We thank the anonymous reviewers and the area chair for their constructive feedback, which helped improve this paper.