Pith. sign in

REVIEW 5 major objections 7 minor 56 references

GADS: A Super Lightweight Model for Head Pose Estimation

T0 review · 5 major / 7 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A 20,000-parameter landmark-only model matches head-pose accuracy of networks thousands of times larger.

desk verdict A plausible 20K-parameter HPE regressor whose efficiency claims hold up, but whose 'competitive accuracy' is not established due to evaluation on FAN-filtered test subsets versus full-set baselines. read the letter →

arxiv 2504.15751 v1 pith:5J5W6KNK submitted 2025-04-22 cs.CV

Menan Velayuthan , Asiri Gawesha , Purushoth Velayuthan , Nuwan Kodagoda , Dharshana Kasthurirathna , Pradeepa Samarasinghe This is my paper · ORCID
classification cs.CV
keywords headposeestimationfaciallandmarksDeepSetsmultiheadattentionlightweightmodeledgecomputingpermutationinvariancegeometriclearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes GADS, a head-pose estimator built from facial landmarks that is much smaller and faster than existing models while staying close to their accuracy. The intended contribution is a deployable baseline for resource-constrained settings: about 20,000 parameters, a reported CPU latency near 2 ms for the landmark-only variant, and mean absolute errors within roughly 0.2 to 1.4 degrees of the best values in the compared tables. The model groups 27 facial landmarks into five regions, reduces each group with a permutation-invariant Deep Set layer, and uses multihead attention to combine the group summaries before regressing yaw, pitch, and roll. A hybrid variant that also consumes the RGB image is evaluated as well. If the claims hold, GADS is the lightest head-pose regression stage reported to date.

What carries the argument

The load-bearing object is the Grouped Attention Deep Set block. A Deep Set layer is a permutation-invariant encoder-aggregator-decoder that maps an unordered set of landmark coordinates to a fixed vector, so the first stage of the model does not depend on landmark ordering. GADS runs five such layers in parallel, one per facial region, aggregates each group with a max operation, and then applies multihead self-attention to the five group summaries so the final regressor sees both within-region geometry and cross-region relationships. This reduces the input from 67x3 coordinates to five 32-dimensional vectors before the final decoder, which is what keeps the parameter count near 20,000.

What would settle it

Run LwPosr, FSA-Net, WHENet, EVA-GCN, and TokenHPE under the exact Protocol 1 pipeline used for GADS, with the same face crops, landmark detector, and test splits, and compare mean absolute errors; if the re-measured baselines are more than about 1.5 degrees better than the borrowed numbers, the competitive-accuracy claim is unsupported. Separately, measure end-to-end CPU latency including face detection and landmark extraction to test the deployment-speed claim.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a pose regressor can be reduced to a twenty-thousand-parameter landmark-only network without leaving the accuracy band of state-of-the-art models. GADS partitions 68 FAN landmarks (27 used) into left eye, right eye, left cheek, right cheek, and chin; each group is processed by a dedicated Deep Set encoder-decoder with a max invariant operator, yielding a fixed-size summary per region. Multihead attention with four heads then mixes these summaries, and a two-layer decoder regresses yaw, pitch, and roll. Trained on 300W-LP with MAE loss, GADS reports a BIWI MAE of 3.90 and an AFLW2000 MAE of 5.30 under Protocol 1, including the best yaw error on BIWI at 3.61, while using 0.02 million parameters; the authors state this is 7.5x smaller than LwPosr and 4321x smaller than TokenHPE, with measured CPU latency of 2.04 ms for the landmark-only model. The hybrid variant, adding a small CNN on the 64x64 RGB crop, reports a BIWI Protocol 2 MAE of 3.46, outperforming vanilla GADS on that split.

Load-bearing premise

The central comparison numbers for competing methods are taken from a previous study's tables rather than re-measured, so the claim that a 20,000-parameter model is competitive depends on those external mean absolute errors having been produced under the same training and evaluation conditions as GADS.

Editorial extensions

If this is right

  • Because GADS is landmark-only and adds about 2 ms and 20,000 parameters, it can be inserted as a near-free head-pose stage into systems that already compute facial landmarks for other tasks.
  • The 0.02-million-parameter size makes the model a candidate for edge and compute-poor deployments where RGB models in the hundreds of thousands or millions of parameters are impractical.
  • On the cross-dataset protocol, GADS reports the best yaw error on BIWI among the compared methods, with pitch and roll slightly behind, suggesting the geometric grouping captures orientation information unevenly across angles.
  • The same grouped Deep Set plus attention design is presented as extensible to any landmark-based analysis, not only head pose estimation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The semantic grouping into eyes, cheeks, and chin may itself be a large part of the inductive bias; a controlled experiment with landmarks randomly permuted into five arbitrary groups would reveal whether the region grouping or merely the set partitioning drives accuracy.
  • The reported 2 ms latency covers only the pose regressor, not face detection or landmark extraction; an end-to-end measurement that includes the upstream detector would show what share of the real-time budget remains on a given device.
  • If the same 20K-parameter budget transfers, the grouped-set design should produce comparable accuracy on other landmark regression problems such as hand pose or gaze estimation, which would be a direct test of the paper's generality claim.
  • Re-running the baseline models under the exact same preprocessing pipeline would refine the competitive-accuracy claim, since the comparison currently depends on numbers taken from earlier studies.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 7 minor

Summary. This paper introduces GADS, a landmark-only head-pose estimator with about 20K parameters, together with a hybrid variant (GADS-Hybrid) that additionally consumes a 64x64 RGB crop. The vanilla model applies five parallel Deep Set layers to region-wise groups of 3D facial landmarks, passes the group embeddings through a four-head self-attention layer, and regresses yaw, pitch, and roll with a small MLP. The authors train on 300W-LP and on a 70% split of BIWI, evaluate under two protocols on AFLW2000 and BIWI, report MAE values, an ablation over architecture choices, and CPU/GPU execution times. The central claims are that GADS is the smallest published HPE model (about 7.5x smaller than LwPosr and 4321x smaller than TokenHPE), executes tens of times faster than lightweight CNN baselines, and achieves accuracy competitive with state-of-the-art methods.

Significance. The strongest assets of the paper are its genuinely tiny architecture and the clarity of the size and single-forward-pass timing measurements: a 20K-parameter regressor with sub-3ms per-forward latency is a plausible deployable component for landmark-based pipelines, and the Deep Set plus grouped-attention design is simple and extensible. The ablation study is broad, and the authors are transparent about the MTCNN/FAN dependency. If the accuracy comparison were re-run on matched evaluation conditions, the 'competitive with SOTA' claim would be a useful data point for the HPE community. As it stands, however, the evidence for that claim is not yet reliable: the SOTA numbers in Tables 2-3 are copied from a single prior study, GADS is scored on FAN-filtered easy subsets rather than the full test sets, and hyperparameters were selected using the test partitions themselves. The paper provides no code and re-runs no baselines, so the quantitative claims are not independently verifiable at this stage.

major comments (5)
  1. [Section 3.3, Table 1, Tables 2-3] The headline accuracy comparison is not apples-to-apples. Table 1 shows that GADS is evaluated only on frames where FAN landmark detection succeeded (BIWI 14,954/15,678; 300W-LP 105,484/122,450; AFLW2000 1,869/2,000), while the SOTA values in Tables 2 and 3 are copied from the LwPosr study (per the captions) and correspond to the full test sets. Section 5.5 states that FAN fails on occlusions, warped images, and extreme angles, and that pose error is largest at high angles. The comparison therefore omits exactly the hard cases for the competing methods while evaluating GADS only on the easy cases, which biases the comparison in GADS's favor. This is a concrete comparability defect, not merely a difference in training recipes. The authors must re-run or re-evaluate all baselines on the identical surviving frame subsets, or evaluate GADS on the full test sets with a fallback mechanism, and report both sets of numbers.
  2. [Section 5.4, Table 5, Section 4.0.3] The model-selection procedure leaks the test sets into the reported numbers. The ablation study is conducted under Protocol 1, i.e., with BIWI and AFLW2000 as evaluation sets, and Section 4.0.3 states that all architecture hyperparameters were determined from this ablation. The configuration so selected is then the one whose MAE values are reported in Tables 2-3. Since no separate validation split is described, the headline accuracies are the result of selection on the test partitions themselves and are optimistically biased. Please redo the hyperparameter selection on a held-out validation split (for example, a subset of the training data) and report the final test-set numbers for the single configuration chosen without test-set feedback.
  3. [Section 5.1.2, Table 3; Section 5.1.1, Table 2] Several textual ranking claims are contradicted by the paper's own tables. In Protocol 2, GADS-Hybrid has MAE 3.46 and ranks 6th behind TokenHPE (2.49), 6DRepNet (2.66), TriNet (2.80), Martin (2.90), and VGG16-RNN (3.07), so the claim that it 'secures a position within the top 3 among SOTA models' is unsupported. In Protocol 1 on AFLW2000, GADS's pitch (7.06) ranks 13th of 18 and its roll (5.00) ranks 10th of 18, contradicting the statement that GADS 'ranks within the top 5' in pitch and roll. The stated roll margin of 0.18 in Section 5.1.1 also does not match Table 2 (GADS 3.04 versus TokenHPE 2.71 is a difference of 0.33). All ranking and margin statements must be recomputed from the tables and corrected.
  4. [Abstract, Section 5.2, Table 4, Section 5.5] The speed claims are misattributed and measure only the regressor, not the full pipeline. The abstract says GADS is '25x faster than the current lightest state-of-the-art model', which the paper identifies as LwPosr (Section 2.1), but Table 4 reports execution times only for HopeNet and FSA-Net; no LwPosr timing is given, so that specific claim is unsubstantiated. Section 5.5 says GADS exhibits a processing speed '2x faster than FSANET', which contradicts Table 4 (about 44x faster on CPU and 39x faster on GPU). In addition, the GADS timing of 2.04 ms is for the model alone ('for a given set of landmarks', Section 5.2), whereas HopeNet and FSA-Net are end-to-end models; the MTCNN face detector and FAN landmark extractor add latency, as Section 5.5 itself acknowledges. Please report end-to-end pipeline latency for all methods and correct the inconsistent speedup factors.
  5. [Section 3.2, Section 5.5] The hybrid model's stated purpose is robustness ('To make GADS robust', Section 3.2), but no experiment evaluates the model on the failure cases identified in Section 5.5 (occlusions, warped images, extreme angles). Those cases are precisely the frames excluded from Tables 2-3 by FAN failures, so the paper never demonstrates that GADS-Hybrid recovers accuracy where vanilla GADS fails. An evaluation on the excluded frames or on a hard-angle subset would directly test the robustness motivation and should be added.
minor comments (7)
  1. [Section 3.6] The subsection 'Implementation Details' is empty; the relevant content appears in Section 4.0.3. The empty subsection should be removed or filled.
  2. [Table 1] The AFLW2000 row is misformatted as 'AFLW 2000 1869 6.55', which can be misread as a dataset name followed by only two numbers; please write 'AFLW2000 2000 1869 6.55' so the Before, After, and Loss columns are unambiguous.
  3. [Section 3.4] The Protocol P2 paragraph ends with the stray words 'improve this', which appears to be a leftover editing note and should be removed.
  4. [Section 5.1.2] The model name 'LowPosr' is a typo for 'LwPosr'; please correct it for consistency with Tables 2-3 and the references.
  5. [Section 2.3 and References] The first reference entry is incomplete: it reads 'Self-Attention with Relative Position Representations' with no author names and an in-text citation key '(sha, 2018)' that is not a standard author-year key. Please complete the entry and fix the citation.
  6. [Figures 7 and 8] The captions read 'AFLW200' instead of 'AFLW2000', and the text contains frequent typographical errors such as 'it's' for 'its' and 'di fferent' for 'different'; a careful proofreading pass is needed.
  7. [Introduction and Tables 2-3] Since GADS is presented as an improvement over the authors' prior DS-HPE model, including DS-HPE's MAE, parameter count, and execution time in Tables 2-4 would substantiate the claimed improvement and better situate the contribution; this baseline is currently absent from all comparison tables.

Circularity Check

1 steps flagged · score 6.0 of 10

Final accuracy claim is partly a selection artifact: hyperparameters are tuned on the exact Protocol 1 test sets (BIWI and AFLW2000) whose MAE is then reported as the headline result.

  1. fitted input called prediction [Sections 3.4, 4.0.3, and 5.4; Tables 2 and 5]
    "All hyperparameters of the model architectures were determined based on the results of the ablation study ( Section 5.4). ... To assess the performance of our proposed method under varying architecture parameters, we conducted an ablation study using Protocol 1 ( Table 5). ... The evaluation is performed on the BIWI dataset (Fanelli et al., 2013) and the AFLW2000 dataset (Zhu et al., 2016). ... We employed these best hyperparameters obtained through the ablation study for both the vanilla GADS architecture and the GADS hybrid model architecture."

    Protocol 1 defines the exact test sets on which Table 2 reports the final MAE. The ablation study selected the number of heads, decoder/encoder layers, activation, loss function, learning rate, and dropout by minimizing MAE on those same BIWI and AFLW2000 test sets, and then the GADS row in Table 2 is precisely the chosen configuration's MAE on those same sets. No held-out validation set is described. Consequently the headline competitive-accuracy claim is not an independent prediction: the reported numbers are the optimization target evaluated on the same data, so the accuracy comparison is at least partially forced by the selection procedure. The parameter-count and latency claims remain independent, but the central accuracy evidence reduces to fitting on the evaluation benchmark.

full rationale

The architecture definition, parameter counts, and latency measurements are derived from the model itself and are not circular. The self-citations to the authors' prior DS-HPE work are not load-bearing: the core building blocks (Deep Sets, multihead attention) are external prior work, and the paper's contribution is an architectural combination rather than a theorem resting on those self-citations. The one substantive circular step is that all hyperparameters were chosen by ablation on Protocol 1, and Protocol 1 is also the benchmark whose MAE is reported in the headline comparison. Thus the final accuracy is the output of model selection on the test set, not a held-out prediction. Separately, Table 1 shows GADS is evaluated only on FAN-surviving frames (e.g., 1,869 of 2,000 AFLW2000 images) while Table 2's baseline numbers are taken from LwPosr and are not re-run on the same subsets; this is a real benchmark-comparability concern, but it is a fairness/correctness issue rather than a definitional circularity. Overall, the central 'competitive accuracy' claim is partially circular, warranting a score of 6.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. Its central claims rest on the chosen landmark grouping, the tuned hyperparameters, and the trust placed in externally sourced SOTA baselines. The efficiency claims are less assumption-heavy than the accuracy claims.

free parameters (2)
  • Facial landmark grouping choice = 5 groups, 27 of 68 landmarks: left eye 6, right eye 6, left cheek 5, right cheek 5, chin 5
    The partition of the 68 FAN landmarks into five regions and the decision to drop 41 landmarks is made by hand (Section 3, Figure 5). The paper does not test other groupings, so the reported accuracy depends on this unoptimized but influential design decision.
  • Architecture hyperparameters = attention heads=4, 1 encoder and 1 decoder layer in each Deep Set block, 2 decoder layers before output, learning…
    These were chosen in the ablation study (Section 5.4) evaluated on the same Protocol 1 test sets that later produce the headline results. They are therefore tuned to the evaluation benchmarks rather than fixed a priori.
assumptions (4)
  • domain assumption The FAN landmark detector produces sufficiently accurate 3D landmarks on all evaluation datasets
    GADS consumes FAN landmarks exclusively; Table 1 reports 6.55% to 13.86% frame loss through the detector, and any landmark error propagates directly into the pose estimate.
  • domain assumption The SOTA MAE values in Tables 2 and 3, taken from the LwPosr study, are valid and directly comparable to the authors' runs
    No baseline is re-executed; different input modalities, preprocessing pipelines, and training details are ignored when declaring GADS competitive with SOTA.
  • ad hoc to paper Min-max normalization (Eq. 2) with the nose landmark as origin and region-based grouping preserves all pose-relevant information
    The preprocessing is asserted and applied uniformly; the paper does not compare against using all 68 landmarks, different reference points, or learned grouping.
  • domain assumption The evaluation protocols P1 and P2 match those of prior work, making cross-paper comparisons valid
    Protocol P1 trains on 300W-LP and tests on BIWI and AFLW2000; protocol P2 trains on a 70:30 BIWI split. Consistency with FSANet, HopeNet, and LwPosr is claimed but not verified by reruns.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GADS: A Super Lightweight Model for Head Pose Estimation." pith.science (2026). https://pith.science/paper/5J5W6KNK

@misc{pith2026250415751,
  author       = {Pith},
  title        = {Pith review of: GADS: A Super Lightweight Model for Head Pose Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5J5W6KNK}},
  note         = {Machine review of arXiv:2504.15751}
}
abstract

In human-computer interaction, head pose estimation profoundly influences application functionality. Although utilizing facial landmarks is valuable for this purpose, existing landmark-based methods prioritize precision over simplicity and model size, limiting their deployment on edge devices and in compute-poor environments. To bridge this gap, we propose \textbf{Grouped Attention Deep Sets (GADS)}, a novel architecture based on the Deep Set framework. By grouping landmarks into regions and employing small Deep Set layers, we reduce computational complexity. Our multihead attention mechanism extracts and combines inter-group information, resulting in a model that is $7.5\times$ smaller and executes $25\times$ faster than the current lightest state-of-the-art model. Notably, our method achieves an impressive reduction, being $4321\times$ smaller than the best-performing model. We introduce vanilla GADS and Hybrid-GADS (landmarks + RGB) and evaluate our models on three benchmark datasets -- AFLW2000, BIWI, and 300W-LP. We envision our architecture as a robust baseline for resource-constrained head pose estimation methods.

Figures

Figures reproduced from arXiv: 2504.15751 by the authors.

Figure 1
Figure 1. Architectures of the GADS and the Deep Set Layer [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. The GADS Hybrid CNN consists of three convolution blocks with [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Examples of 68 landmarks extracted using the FAN landmark detector on the AFLW2000 dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Illustration of 27 selected landmarks grouped into five sections, each represented by a distinct color. Landmarks within the same color belong to the same [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Illustration of yaw, pitch, and roll angles for 5 instances of the AFLW200 dataset. The first row illustrates the correct or ground-truth angles, the second [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Mean Absolute error (MAE) of Euler angles variation across 100 samples of the three bench-marking data-sets. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Variations in Yaw, Pitch, and Roll Euler angles across 100 samples from the AFLW2000 dataset. The graph illustrates the true values, along with [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Variation in Yaw, Pitch, and Roll Euler angles across 100 samples from the BIWI dataset. The graph highlights the di [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Variation in Yaw, Pitch, and Roll Euler angles across 100 samples from the 300W-LP dataset. The graph illustrates precise predictions for all three [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 24 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    title Self-Attention with Relative Position Representations , publisher Association for Computational Linguistics , address New Orleans, Louisiana

    , year 2018 . title Self-Attention with Relative Position Representations , publisher Association for Computational Linguistics , address New Orleans, Louisiana . https://aclanthology.org/N18-2074, :10.18653/v1/N18-2074

    doi:10.18653/v1/n18-2074 2018

  3. [3]

    , author Chen, X

    author Albiero, V. , author Chen, X. , author Yin, X. , author Pang, G. , author Hassner, T. , year 2021 . title img2pose: Face alignment and detection via 6dof, face pose estimation , in: booktitle 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. pages 7613--7623 . :10.1109/CVPR46437.2021.00753

    arXiv 2021

  4. [4]

    , author Cho, K

    author Bahdanau, D. , author Cho, K. , author Bengio, Y. , year 2014 . title Neural machine translation by jointly learning to align and translate . journal arXiv preprint arXiv:1409.0473

    arXiv 2014

  5. [5]

    , author Jacobs, D.W

    author Belhumeur, P.N. , author Jacobs, D.W. , author Kriegman, D.J. , author Kumar, N. , year 2011 . title Localizing parts of faces using a consensus of exemplars , in: booktitle CVPR 2011 , pp. pages 545--552 . :10.1109/CVPR.2011.5995602

    arXiv 2011

  6. [6]

    , author Fabbri, M

    author Borghi, G. , author Fabbri, M. , author Vezzani, R. , author Calderara, S. , author Cucchiara, R. , year 2020 . title Face-from-depth for head pose estimation on depth images . journal IEEE Transactions on Pattern Analysis and Machine Intelligence volume 42 , pages 596--609 . :10.1109/TPAMI.2018.2885472

    arXiv 2020

  7. [7]

    , author Venturelli, M

    author Borghi, G. , author Venturelli, M. , author Vezzani, R. , author Cucchiara, R. , year 2017 . title Poseidon: Face-from-depth for driver pose estimation , in: booktitle 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pp. pages 5494--5503 . :10.1109/CVPR.2017.583

    work page doi:10.1109/cvpr.2017.583 2017

  8. [8]

    , author Tzimiropoulos, G

    author Bulat, A. , author Tzimiropoulos, G. , year 2017 . title How far are we from solving the 2d i& 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks) , in: booktitle 2017 IEEE International Conference on Computer Vision (ICCV) , pp. pages 1021--1030 . :10.1109/ICCV.2017.116

    doi:10.1109/iccv.2017.116 2017

Show all 56 references
  1. [9]

    , author Chu, Z

    author Cao, Z. , author Chu, Z. , author Liu, D. , author Chen, Y. , year 2020 . title A vector-based representation to enhance head pose estimation . https://arxiv.org/abs/2010.07184, :10.48550/ARXIV.2010.07184

    2020 doi arXiv

  2. [10]

    , author Su, H

    author Charles, R.Q. , author Su, H. , author Kaichun, M. , author Guibas, L.J. , year 2017 . title Pointnet: Deep learning on point sets for 3d classification and segmentation , in: booktitle 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pp. pages 7...

    2017 doi

  3. [11]

    , author Bahdanau, D

    author Chorowski, J.K. , author Bahdanau, D. , author Serdyuk, D. , author Cho, K. , author Bengio, Y. , year 2015 . title Attention-based models for speech recognition . journal Advances in neural information processing systems volume 28

    2015

  4. [12]

    , year 2022

    author Dhingra, N. , year 2022 . title Lwposr: Lightweight efficient fine grained head pose estimation , in: booktitle 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , publisher IEEE Computer Society , address Los Alamitos, CA, USA . pp. pages 1204--...

    2022

  5. [13]

    , author Beyer, L

    author Dosovitskiy, A. , author Beyer, L. , author Kolesnikov, A. , author Weissenborn, D. , author Zhai, X. , author Unterthiner, T. , author Dehghani, M. , author Minderer, M. , author Heigold, G. , author Gelly, S. , et al., year 2020 . title An image is worth 16x16 words: ...

    2020 arXiv

  6. [14]

    , author Dantone, M

    author Fanelli, G. , author Dantone, M. , author Gall, J. , author Fossati, A. , author Van Gool, L. , year 2013 . title Random forests for real time 3d face analysis . journal International journal of computer vision volume 101 , pages 437--458

    2013

  7. [15]

    , author Wang, Z

    author Gu, J. , author Wang, Z. , author Kuen, J. , author Ma, L. , author Shahroudy, A. , author Shuai, B. , author Liu, T. , author Wang, X. , author Wang, G. , author Cai, J. , et al., year 2018 . title Recent advances in convolutional neural networks . journal Pattern reco...

    2018

  8. [16]

    , author Yang, X

    author Gu, J. , author Yang, X. , author De Mello, S. , author Kautz, J. , year 2017 . title Dynamic facial analysis: From bayesian filtering to recurrent neural network , in: booktitle 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pp. pages 1531--15...

    2017 doi

  9. [17]

    , author Wang, H

    author Guo, Y. , author Wang, H. , author Hu, Q. , author Liu, H. , author Liu, L. , author Bennamoun, M. , year 2021 . title Deep learning for 3d point clouds: A survey . journal IEEE Transactions on Pattern Analysis and Machine Intelligence volume 43 , pages 4338--4364 . :10...

    2021

  10. [18]

    , author Abdelrahman, A.A

    author Hempel, T. , author Abdelrahman, A.A. , author Al-Hamadi, A. , year 2022 . title 6d rotation representation for unconstrained head pose estimation , in: booktitle 2022 IEEE International Conference on Image Processing (ICIP) , pp. pages 2496--2500 . :10.1109/ICIP46576.2...

    2022

  11. [19]

    , author Sullivan, J

    author Kazemi, V. , author Sullivan, J. , year 2014 . title One millisecond face alignment with an ensemble of regression trees , in: booktitle Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition , pp. pages 1867--1874 . :10.1109/CVPR...

    2014 doi

  12. [20]

    , author Alavi, A

    author Kumar, A. , author Alavi, A. , author Chellappa, R. , year 2017 . title Kepler: Keypoint and pose estimation of unconstrained faces by learning efficient h-cnn regressors , in: booktitle 2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG...

    2017 doi

  13. [21]

    , author Wohlhart, P

    author Köstinger, M. , author Wohlhart, P. , author Roth, P.M. , author Bischof, H. , year 2011 . title Annotated facial landmarks in the wild: A large-scale, real-world database for facial landmark localization , in: booktitle 2011 IEEE International Conference on Computer Vi...

    2011

  14. [22]

    , author Juge, R

    author Lathuilière, S. , author Juge, R. , author Mesejo, P. , author Muñoz-Salinas, R. , author Horaud, R. , year 2017 . title Deep mixture of linear inverse regressions applied to head-pose estimation , in: booktitle 2017 IEEE Conference on Computer Vision and Pattern Recogn...

    2017 doi

  15. [23]

    , author Bottou, L

    author Lecun, Y. , author Bottou, L. , author Bengio, Y. , author Haffner, P. , year 1998 . title Gradient-based learning applied to document recognition . journal Proceedings of the IEEE volume 86 , pages 2278--2324 . :10.1109/5.726791

    1998 doi

  16. [24]

    , author Chen, Y

    author Liu, H. , author Chen, Y. , author Zhao, W. , author Zhang, S. , author Zhang, Z. , year 2021 a. title Human pose recognition via adaptive distribution encoding for action perception in the self-regulated learning process . journal Infrared Physics I& Technology volume ...

    2021

  17. [25]

    , author Fang, S

    author Liu, H. , author Fang, S. , author Zhang, Z. , author Li, D. , author Lin, K. , author Wang, J. , year 2022 a. title Mfdnet: Collaborative poses perception and matrix fisher distribution for head pose estimation . journal IEEE Transactions on Multimedia volume 24 , page...

    2022

  18. [26]

    , author Liu, T

    author Liu, H. , author Liu, T. , author Chen, Y. , author Zhang, Z. , author Li, Y.F. , year 2022 b. title Ehpe: Skeleton cues-based gaussian coordinate encoding for efficient human pose estimation . journal IEEE Transactions on Multimedia , pages 1--12 :10.1109/TMM.2022.3197364

    2022

  19. [27]

    , author Liu, T

    author Liu, H. , author Liu, T. , author Zhang, Z. , author Sangaiah, A.K. , author Yang, B. , author Li, Y. , year 2022 c. title Arhpe: Asymmetric relation-aware representation learning for head pose estimation in industrial human–computer interaction . journal IEEE Transacti...

    2022

  20. [28]

    , author Nie, H

    author Liu, H. , author Nie, H. , author Zhang, Z. , author Li, Y.F. , year 2021 b. title Anisotropic angle distribution learning for head pose estimation and attention understanding in human-computer interaction . journal Neurocomputing volume 433 , pages 310--322 . https://w...

    2021 doi

  21. [29]

    , author Zheng, C

    author Liu, H. , author Zheng, C. , author Li, D. , author Shen, X. , author Lin, K. , author Wang, J. , author Zhang, Z. , author Zhang, Z. , author Xiong, N.N. , year 2022 d. title Edmf: Efficient deep matrix factorization with review feature learning for industrial recommen...

    2022

  22. [30]

    , author Wang, J

    author Liu, T. , author Wang, J. , author Yang, B. , author Wang, X. , year 2021 c. title Ngdnet: Nonuniform gaussian-label distribution learning for infrared head pose estimation and on-task behavior understanding in the classroom . journal Neurocomputing volume 436 , pages 2...

    2021 doi

  23. [31]

    , author Tang, J

    author Lugaresi, C. , author Tang, J. , author Nash, H. , author McClanahan, C. , author Uboweja, E. , author Hays, M. , author Zhang, F. , author Chang, C. , author Yong, M.G. , author Lee, J. , author Chang, W. , author Hua, W. , author Georg, M. , author Grundmann, M. , yea...

    2019 arXiv

  24. [32]

    , author Pham, H

    author Luong, T. , author Pham, H. , author Manning, C.D. , year 2015 . title Effective approaches to attention-based neural machine translation , in: editor M \`a rquez, L. , editor Callison-Burch, C. , editor Su, J. (Eds.), booktitle Proceedings of the 2015 Conference on Emp...

    2015 doi

  25. [33]

    , author Van De Camp, F

    author Martin, M. , author Van De Camp, F. , author Stiefelhagen, R. , year 2014 . title Real time head model creation and head pose estimation on consumer depth cameras , in: booktitle 2014 2nd International Conference on 3D Vision , pp. pages 641--648 . :10.1109/3DV.2014.54

    2014 doi

  26. [34]

    , author Gawesha, A

    author Menan, V. , author Gawesha, A. , author Samarasinghe, P. , author Kasthurirathna, D. , year 2023 . title Ds-hpe: Deep set for head pose estimation , in: booktitle 2023 IEEE 13th Annual Computing and Communication Workshop and Conference (CCWC) , pp. pages 1179--1184 . :...

    2023

  27. [35]

    , author Robertson, N.M

    author Mukherjee, S.S. , author Robertson, N.M. , year 2015 . title Deep head pose: Gaze-direction estimation in multimodal video . journal IEEE Transactions on Multimedia volume 17 , pages 2094--2107 . :10.1109/TMM.2015.2482819

    2015

  28. [36]

    , author Yi, L

    author Qi, C.R. , author Yi, L. , author Su, H. , author Guibas, L.J. , year 2017 . title Pointnet++: Deep hierarchical feature learning on point sets in a metric space . journal Advances in neural information processing systems volume 30

    2017

  29. [37]

    , author De Mello, S

    author Ranjan, R. , author De Mello, S. , author Kautz, J. , year 2018 . title Light-weight head pose invariant gaze tracking , in: booktitle 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pp. pages 2237--22378 . :10.1109/CVPRW.2018.00290

    2018

  30. [38]

    , author Chong, E

    author Ruiz, N. , author Chong, E. , author Rehg, J.M. , year 2018 . title Fine-grained head pose estimation without keypoints , in: booktitle Proceedings of the IEEE conference on computer vision and pattern recognition workshops , pp. pages 2074--2083

    2018

  31. [39]

    , author Tzimiropoulos, G

    author Sagonas, C. , author Tzimiropoulos, G. , author Zafeiriou, S. , author Pantic, M. , year 2013 . title 300 faces in-the-wild challenge: The first facial landmark localization challenge , in: booktitle 2013 IEEE International Conference on Computer Vision Workshops , pp. ...

    2013 doi

  32. [40]

    , author Sun, Z

    author Shao, M. , author Sun, Z. , author Ozay, M. , author Okatani, T. , year 2019 . title Improving head pose estimation with a combined loss and bounding box margin adjustment . https://arxiv.org/abs/1905.08609, :10.48550/ARXIV.1905.08609

    2019 doi arXiv

  33. [41]

    , author Shazeer, N

    author Vaswani, A. , author Shazeer, N. , author Parmar, N. , author Uszkoreit, J. , author Jones, L. , author Gomez, A.N. , author Kaiser, . , author Polosukhin, I. , year 2017 . title Attention is all you need . journal Advances in neural information processing systems volume 30

    2017

  34. [42]

    , author Alameda-Pineda, X

    author Wang, W. , author Alameda-Pineda, X. , author Xu, D. , author Fua, P. , author Ricci, E. , author Sebe, N. , year 2018 . title Every smile is unique: Landmark-guided diverse smile generation , in: booktitle Proceedings of the IEEE Conference on Computer Vision and Patte...

    2018

  35. [43]

    , author Samarasinghe, P

    author Wedasingha, N. , author Samarasinghe, P. , author Seneviratne, L. , author Puiatti, A. , author Papandrea, M. , author Dhanayaka, D. , year 2022 . title Skeleton based periodicity analysis of repetitive actions , in: booktitle 2022 IEEE 7th International conference for ...

    2022

  36. [44]

    , author Sang, N

    author Xiao, S. , author Sang, N. , author Wang, X. , author Ma, X. , year 2020 . title Leveraging ordinal regression with soft labels for 3d head pose estimation from point sets , in: booktitle ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal P...

    2020

  37. [45]

    , author Mo, S

    author Xin, M. , author Mo, S. , author Lin, Y. , year 2021 . title Eva-gcn: Head pose estimation based on graph convolutional networks , in: booktitle 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pp. pages 1462--1471 . :10.1109/CVPRW...

    2021

  38. [46]

    , author Chen, Y.T

    author Yang, T.Y. , author Chen, Y.T. , author Lin, Y.Y. , author Chuang, Y.Y. , year 2019 . title Fsa-net: Learning fine-grained structure aggregation for head pose estimation from a single image , in: booktitle 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2019

  39. [47]

    , author Huang, Y.H

    author Yang, T.Y. , author Huang, Y.H. , author Lin, Y.Y. , author Hsiu, P.C. , author Chuang, Y.Y. , year 2018 . title SSR-Net: A Compact Soft Stagewise Regression Network for Age Estimation , in: booktitle Proceedings of the Twenty-Seventh International Joint Conference on A...

    2018 doi

  40. [48]

    , author Kottur, S

    author Zaheer, M. , author Kottur, S. , author Ravanbakhsh, S. , author P \' o czos, B. , author Salakhutdinov, R. , author Smola, A.J. , year 2017 . title Deep sets . journal CoRR volume abs/1703.06114 . http://arxiv.org/abs/1703.06114, http://arxiv.org/abs/1703.06114 arXiv:1...

    2017 arXiv

  41. [49]

    , author Liu, H

    author Zhang, C. , author Liu, H. , author Deng, Y. , author Xie, B. , author Li, Y. , year 2023 . title Tokenhpe: Learning orientation tokens for efficient head pose estimation via transformers , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Vision and Patt...

    2023

  42. [50]

    , author Zhang, Z

    author Zhang, K. , author Zhang, Z. , author Li, Z. , author Qiao, Y. , year 2016 . title Joint face detection and alignment using multitask cascaded convolutional networks . journal IEEE Signal Processing Letters volume 23 , pages 1499--1503 . https://doi.org/10.1109

    2016

  43. [51]

    , author Fan, H

    author Zhou, E. , author Fan, H. , author Cao, Z. , author Jiang, Y. , author Yin, Q. , year 2013 . title Extensive facial landmark localization with coarse-to-fine convolutional network cascade , in: booktitle 2013 IEEE International Conference on Computer Vision Workshops , ...

    2013 doi

  44. [52]

    , author Gregson, J

    author Zhou, Y. , author Gregson, J. , year 2020 . title Whenet: Real-time fine-grained estimation for wide range head pose . https://arxiv.org/abs/2005.10353, :10.48550/ARXIV.2005.10353

    2020 doi arXiv

  45. [53]

    , author Lei, Z

    author Zhu, X. , author Lei, Z. , author Liu, X. , author Shi, H. , author Li, S.Z. , year 2016 . title Face alignment across large poses: A 3d solution , in: booktitle Proceedings of the IEEE conference on computer vision and pattern recognition , pp. pages 146--155

    2016

  46. [54]

    , author Liu, X

    author Zhu, X. , author Liu, X. , author Lei, Z. , author Li, S.Z. , year 2019 . title Face alignment in full pose range: A 3d total solution . journal IEEE Transactions on Pattern Analysis and Machine Intelligence volume 41 , pages 78--92 . https://doi.org/10.1109

    2019

  47. [55]

    , author Ramanan, D

    author Zhu, X. , author Ramanan, D. , year 2012 . title Face detection, pose estimation, and landmark localization in the wild , in: booktitle 2012 IEEE Conference on Computer Vision and Pattern Recognition , pp. pages 2879--2886 . :10.1109/CVPR.2012.6248014

    2012

  48. [56]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.