Learning to Train with Synthetic Humans

David T. Hoffmann; Dimitrios Tzionas; Micheal J. Black; Siyu Tang

REVIEW 2 major objections 5 minor 18 references

Learning to Train with Synthetic Humans

T0 review · 2 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A teacher network that feeds a pose estimator its hardest synthetic images beats random sampling.

desk verdict Useful synthetic datasets and a clean occlusion-masking result, but the teacher's 0.5 mAP gain may just be static rebalancing; the paper deserves a serious referee with a needed control. read the letter →

arxiv 1908.00967 v1 pith:FPH75C5S submitted 2019-08-02 cs.CV

David T. Hoffmann , Dimitrios Tzionas , Micheal J. Black , Siyu Tang This is my paper · ORCID

classification cs.CV

keywords syntheticdatamulti-personposeestimationstudent-teachertrainingocclusiondomainadaptationhardexampleminingcamerapitchMPIIbenchmark

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

The paper tests whether synthetic humans can substitute for costly manual annotation in multi-person 2D pose estimation, and how to train on them effectively. It builds a purely synthetic dataset (DS), a mixed dataset (DM) that composites synthetic humans onto real MPII training images, and a stylized version (DStyle). Training with either synthetic-data recipe lifts mean average precision (mAP) on MPII from 77.7 to 78.4, but the gain is almost entirely occlusion: when the loss from synthetic humans is masked out, accuracy improves to 78.9 for DM and 79.1 for DStyle. The paper's second claim is that uniform random sampling wastes much of a synthetic dataset; an adversarial teacher that re-weights groups of synthetic images according to the student's current losses improves the purely-synthetic model from 78.4 to 78.9 mAP.

What carries the argument

The load-bearing mechanism is the adversarial student-teacher loop with a reward/penalty update. The student is an OpenPose-style pose network; the teacher is a small convolutional network that takes the student's highest-loss real image from the previous N steps and outputs a probability distribution over 10 groups of synthetic images, grouped by camera pitch or by minimal person-to-person distance. After the student trains on samples from the selected group, the teacher is rewarded if the mean per-joint loss on that synthetic batch is at least the average of the last H losses (Eq. 1); the selected group's probability is then raised and the others lowered (Eq. 2), and the teacher is updated by KL divergence against this pseudo-label. The two dataset variants DS and DM, plus the style-transferred DStyle and the masking of synthetic-human losses, are the experimental apparatus that separates the occlusion effect from the appearance effect.

What would settle it

Run the camera-pitch teacher experiment again but replace the teacher's reward with a random group draw every N steps, keeping all other hyperparameters identical; if mAP stays 78.9, the reward signal is not what drives the gain.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that synthetic humans help multi-person pose estimation mainly as occluders, not as appearance training data, and that a student-teacher loop can mine synthetic data more effectively than random sampling. The evidence for the occlusion claim is the masking ablation: MDR+DM goes from 78.4 to 78.9 mAP when synthetic-human losses are excluded, and MDR+DStyle goes from 78.3 to 79.1, meaning the network is better off learning from the occlusion patterns than from the synthetic humans' pixels. The teacher claim is established with the camera-pitch grouping: an OpenPose-based student trained with real data plus DS under teacher-selected sampling reaches 78.9 mAP, beating the 78.4 of uniform training; the minimal-distance grouping reaches 78.7. The paper also finds that the teacher often converges to uniform sampling over groups, which corresponds to a more extreme camera-pitch distribution than the real training data, and that the largest gains appear for heavily occluded people and uncommon camera positions.

Load-bearing premise

The teacher's entire gain rests on the assumption that a current synthetic-batch loss above a moving average of recent losses reliably flags which of ten pre-defined groups of synthetic images is most useful to the student at that moment.

Editorial extensions

If this is right

Adding synthetic humans to real training images improves multi-person pose estimation on MPII, but only because the synthetic humans create occlusion; their pixels are not useful for learning appearance.
A model trained purely on synthetic data fails badly (13.4 mAP), so synthetic data should be combined with real data or used as occluders rather than relied on alone.
Random sampling of a large synthetic dataset is suboptimal; selecting the currently hardest group of samples, defined by camera pitch or person distance, improves accuracy on highly occluded people and uncommon viewpoints.
The stylized mixed dataset with masked synthetic losses (79.1 mAP) is the overall best model, indicating that closer visual match between synthetic occluders and real scenes helps.
The finding that informative samples change during training suggests that static dataset construction misses value that adaptive curriculum-style sampling can extract.

Reading between the lines

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

Because the masking experiment isolates occlusion, a direct extension is to replace synthetic humans with non-human occluders at matched density; if mAP rises similarly, the benefit is generic occlusion rather than human-specific context.
The teacher's 0.5 mAP gain over uniform sampling is small and rests on one-dimensional binning; a 2D difficulty grid or a learned sampling density over both camera pitch and distance could test whether the mechanism scales.
The paper's comparisons imply that texture diversity, not dataset bias, is the main ceiling on synthetic-data training; improving texture variety should be a cheaper route than more elaborate domain adaptation.
The same teacher could be applied to other detection or segmentation tasks that suffer from occlusion, since nothing in the reward signal is pose-specific beyond the loss definition.

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.

Desk Editor's Note

read the letter

Two things to know. First, the paper's strongest result is not that synthetic data helps — it's that masking the loss on synthetic humans is what helps, and the gain comes from extra occlusion rather than pose supervision. The masking ablation is clean: MDR+DM goes from 78.4 to 78.9 mAP, MDR+DStyle from 78.3 to 79.1. Second, the teacher model is the weakest part. The claimed 0.5 mAP improvement over normal training (78.9 vs 78.4 on DS with camera-pitch grouping) is not actually isolated as an effect of adaptive difficulty selection.

What the paper does well: the datasets are carefully generated and released, the synthetic pipeline is sensible, and the paper is honest about its own caveats. The occlusion-masking result is genuinely new as far as I know, and it's useful to practitioners. The related work is properly situated, and the self-citations in the synthetic-data line are not a problem.

Where the soft spots are: all headline differences are a few tenths of mAP with no error bars or significance tests. That alone would warrant a cautious verdict. But the real issue is the teacher. In Section 5.3 the authors themselves note that the teacher typically converges to a uniform sampling strategy over the groups. A uniform distribution over camera-pitch bins is a different static training distribution than the natural one — the real training data is biased toward low camera pitch. They never compare against a fixed uniform sampler over the same groups. So the 0.5 mAP gain could be entirely due to re-weighting toward rare pitch angles, with the reward/update mechanism in Eqs. 1–2 contributing nothing. The stress-test note about this is exactly right. The authors' own limitations section admits the grouping is one-dimensional and hand-chosen, and the conclusion discloses that their best model is actually MDR+DStyle+masks, not the teacher. None of this invalidates the dataset contribution or the masking result, but the 'student-teacher outperforms normal training' claim is not supported as a statement about adaptive sampling.

Who this is for: people working on synthetic training data for pose estimation, and anyone interested in curriculum learning. The masking insight and the datasets are worth engaging with; the teacher needs a control baseline and proper statistics.

Recommendation for peer review: a serious referee should engage with it. The masking result deserves to be in the literature. The teacher section needs a substantial revision — add a static uniform-group control and error bars, and temper the claim accordingly.

Referee Report

2 major / 5 minor

Summary. The paper introduces two synthetic multi-person pose datasets — a purely synthetic dataset DS and a mixed dataset DM that augments MPII with synthetic humans, plus a stylized version DStyle — and studies their use in training an OpenPose-style network. It reports that training on real plus synthetic data improves held-out MPII mAP over real-only training (78.4 vs 77.7), that masking the loss from synthetic humans further improves accuracy (78.9/79.1), and that an adversarial student-teacher that samples synthetic groups according to current difficulty outperforms normal training on DS (78.9 vs 78.4). The paper concludes that the benefit of DM/DStyle comes mainly from additional occlusion and that informed sampling helps.

Significance. If the results hold, the paper provides a useful analysis of synthetic data for multi-person pose estimation, with two notable assets: a careful held-out MPII evaluation and a clean ablation showing that masking the synthetic-generated loss improves mAP, supporting the occlusion explanation. The release of datasets and code is also a concrete contribution, and the negative result on domain stylization is informative. However, the headline student-teacher result is currently not isolated from a static rebalancing effect, and all improvements are within a few tenths of mAP without error bars or significance tests. The masking result is the more defensible claim; the teacher contribution needs additional controls.

major comments (2)

[Section 5.3, Table 3, Eqs. (1)-(2)] The central claim that the adversarial teacher 'outperforms normal training' is confounded because the teacher is compared only against MDR+DS, which samples groups according to the natural, camera-pitch-biased distribution. The paper states in Section 5.3 that the teacher typically converges to a uniform sampling strategy over the 10 groups. A fixed uniform sampler over the same 10 groups is the necessary control: without it, the 0.5 mAP gain could be entirely due to rebalancing toward rare camera-pitch groups rather than to the reward/update mechanism in Eqs. (1)-(2). Please add this baseline or substantially weaken the claim.
[Tables 2 and 3, Fig. 6] The headline improvements (78.4 to 78.9 for DM+masks, 78.4 to 79.1 for DStyle+masks, and 78.4 to 78.9 for Teacher-C) are all within a few tenths of mAP, and no error bars, multiple seeds, or significance tests are reported. Given that the Fig. 6 caption itself states the teacher methods 'are hard to distinguish,' the improvement claims are not yet supported beyond possible noise. Please report variance across at least a few training runs or otherwise quantify the stability of these differences.

minor comments (5)

[Section 5.3 / Fig. 6] The text says improvements for highly occluded people are strongest for models trained with the teacher, but the Fig. 6 caption says the teacher methods are hard to distinguish; please reconcile these statements.
[Section 5.2] The explanation that style-transfer artifacts cause DStyle to underperform DM is speculative; a quantitative analysis of failure cases (e.g., fraction of ghost-like artifacts) would strengthen the claim.
[Section 3.3] The phrase 'We render each of the 15,956 images in our training set 5 times' is ambiguous about whether 'training set' refers to MPII training images or the DM frames generated from them; please clarify.
[Section 4.1] The choice to feed the teacher the real image with the highest mean loss per joint from the previous N steps is not motivated; please explain why this particular representation of student state is appropriate.
[Section 5.3] The sentence 'This is not equivalent to random sampling, as samples are not uniformly distributed across groups' is confusing because the teacher's converged distribution is described as uniform over groups; what the authors likely mean is that the natural data distribution is not uniform, so sampling uniformly over groups differs from sampling from the natural distribution. Please rephrase.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: teacher gains are judged on held-out MPII validation, and the self-cited assets are not load-bearing for the headline claims.

full rationale

The paper's central comparisons are not circular. All three headline results — synthetic augmentation improving the real-only baseline (Table 1), masking out synthetic-human loss improving accuracy (Table 2, Section 5.2), and the teacher's group-sampling scheme beating natural-distribution sampling on the purely synthetic dataset (Table 3) — are measured on a held-out validation split of MPII that is disjoint from training (Section 5). The teacher's reward/update (Eqs. 1–2) defines "difficult" as high student loss on synthetic data, so the claim that the teacher selects high-loss groups is true by definition, but the downstream claim that this improves real-data mAP is an external, falsifiable result. No equation in the paper is equivalent by construction to a later claim. The heavy use of prior author-built resources (SMPL+H [39], MoSh [26], SURREAL textures [18], embodied hands [13]) supplies data-generation assets, not the target conclusions; these are pre-existing, independently published components rather than results derived in this paper. The stated limitations (grouping is one-dimensional, group size/spacing may matter, teacher tested only on OpenPose/pose estimation) are acknowledged confounds and scope restrictions, not circularity. The absence of a static-uniform-sampling control for the teacher (Section 5.3) is a legitimate experimental weakness — the teacher's convergence to uniform sampling means the 0.5 mAP gain might be due to rebalancing camera pitch rather than to adaptive selection — but that is a missing control and a threat to internal validity, not a derivation that reduces to its own assumptions. Therefore no circular step is exhibited.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claims rest on several hand-set design parameters and domain assumptions. The most consequential are the Poisson counts, the camera pitch range, the teacher groups and reward hyperparameters, and the unstated trust in SMPL+H/MoSh, Mask-RCNN masks, and the loss-as-informativeness proxy. No new physical or conceptual entities are postulated.

free parameters (8)

Poisson lambda for DS = 9
Chosen to encourage many synthetic humans per image while avoiding extreme values (Sec. 3.2).
Poisson lambda for DM = 4
Chosen to add occlusion without overwhelming the real MPII image (Sec. 3.3).
Camera pitch range = [0, 45] degrees
Sampled uniformly; preliminary experiments indicated camera angle affects difficulty (Sec. 3.2).
Number of teacher groups = 10
Assumed trade-off between precision and teacher difficulty (Sec. 4).
Teacher exploration probability epsilon = 0.1
Chosen for exploration/exploitation balance (Sec. 4.1).
Teacher update coefficient alpha = not reported
Controls update size in Eq. 2; value needed to reproduce teacher training and not given in text.
Teacher history length H = not reported
Number of past losses averaged in Eq. 1; not specified.
Group boundaries for min distance and camera pitch = linear spacing with [0,640) px; [min+Var, max-Var)
Hand-defined grouping rule; affects which samples the teacher can select (Sec. 4).

assumptions (6)

domain assumption SMPL+H and MoSh produce plausible human body poses and shapes
The realism of the synthetic humans depends on these prior models and fitting procedures (Sec. 3.1).
domain assumption Mask-RCNN detects humans in background images accurately enough
Used to reject background images containing humans and to create style-transfer masks; errors propagate to the data (Sec. 3.2, Sec. 3.4).
domain assumption Student synthetic-data loss is a valid proxy for sample informativeness
The teacher reward in Eq. 1 assumes higher student loss means a more useful training sample.
ad hoc to paper Camera pitch and minimal person distance capture image difficulty
The grouping is based on these two scalar features, justified only by 'we found empirically' (Sec. 4).
domain assumption Freezing the first four feature extractor layers when training on synthetic data is beneficial
Adopted from Hinterstoisser et al. [18] without an ablation on this dataset.
domain assumption Occlusion by synthetic humans approximates realistic occlusion
The mixed dataset design and the occlusion-based conclusion assume synthetic occluders transfer to real scenes (Sec. 3.3, Sec. 5.2).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning to Train with Synthetic Humans." pith.science (2026). https://pith.science/paper/FPH75C5S

@misc{pith2026190800967,
  author       = {Pith},
  title        = {Pith review of: Learning to Train with Synthetic Humans},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FPH75C5S}},
  note         = {Machine review of arXiv:1908.00967}
}

read the original abstract

Neural networks need big annotated datasets for training. However, manual annotation can be too expensive or even unfeasible for certain tasks, like multi-person 2D pose estimation with severe occlusions. A remedy for this is synthetic data with perfect ground truth. Here we explore two variations of synthetic data for this challenging problem; a dataset with purely synthetic humans and a real dataset augmented with synthetic humans. We then study which approach better generalizes to real data, as well as the influence of virtual humans in the training loss. Using the augmented dataset, without considering synthetic humans in the loss, leads to the best results. We observe that not all synthetic samples are equally informative for training, while the informative samples are different for each training stage. To exploit this observation, we employ an adversarial student-teacher framework; the teacher improves the student by providing the hardest samples for its current state as a challenge. Experiments show that the student-teacher framework outperforms normal training on the purely synthetic dataset.

Figures

Figures reproduced from arXiv: 1908.00967 by the authors.

**Figure 1.** Qualitative comparisons between our models and Cao et al. [8]. Cao et al. Ours (using teacher) Ours (best) Cao et al. Ours (best) Cao et al. Ours (best) A B We generate both types of synthetic datasets. One purely synthetic dataset and a mixed dataset, which is generated by augmenting the MPII pose estimation dataset [3] with synthetic humans. In particular, we design these datasets to improve on frequent failure c… view at source ↗

**Figure 2.** Schematic of the data generation pipeline. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Example images from the purely synthetic dataset. It contains high occlu [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: (A) Example images from DM. (B) Corresponding images of DStyle. For the last image, the segmentation network included non-human parts in the segmentation masks. Resulting artifacts can be seen for rightmost synthetic human. Mixed Mixed Mixed Stylized Stylized Stylized…

**Figure 5.** Figure 5: Diagram of the forward pass. The total loss is denoted as [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Detection performance for varying ratio of visible joints (mAP). The [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 17 canonical work pages

[1]

http://mocap.cs.cmu.e du

Carnegie-mellon mocap database. http://mocap.cs.cmu.e du

work page
[2]

https://github.com/matterport/Mask RCNN (2017)

Abdulla, W.: Mask r-cnn for object detection and instance seg mentation on keras and tensorﬂow. https://github.com/matterport/Mask RCNN (2017)

work page 2017
[3]

In: Proceedings of the IEEE Conference on Compute r Vision and Pattern Recognition

Akhter, I., Black, M.J.: Pose-conditioned joint angle limi ts for 3d human pose reconstruction. In: Proceedings of the IEEE Conference on Compute r Vision and Pattern Recognition. pp. 1446–1455 (2015)

work page 2015
[4]

In: 2017 IEEE Conference on Comput er Vision and Pattern Recognition (CVPR)

Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi- person 2d pose estima- tion using part aﬃnity ﬁelds. In: 2017 IEEE Conference on Comput er Vision and Pattern Recognition (CVPR). pp. 1302–1310. IEEE (2017)

work page 2017
[5]

In: Pro- ceedings of the European Conference on Computer Vision (ECCV)

Fabbri, M., Lanzi, F., Calderara, S., Palazzi, A., Vezzani, R ., Cucchiara, R.: Learn- ing to detect and track visible and occluded body joints in a virt ual world. In: Pro- ceedings of the European Conference on Computer Vision (ECCV). pp. 430–446 (2018)

work page 2018
[6]

Archives of the Game Developers Conference (Mar 2003), http://www.research.scea.com/gdc2003/spherical-harmonic-lighting.pdf

Green, R.: Spherical Harmonic Lighting: The Gritty Details. Archives of the Game Developers Conference (Mar 2003), http://www.research.scea.com/gdc2003/spherical-harmonic-lighting.pdf

work page 2003
[7]

In: Computer Vision (ICCV), 2017 IEEE International Conference on

He, K., Gkioxari, G., Doll´ ar, P., Girshick, R.: Mask r-cnn. In: Computer Vision (ICCV), 2017 IEEE International Conference on. pp. 2980–2988. I EEE (2017)

work page 2017
[8]

In: International Conference on Learning Representations (ICLR)

Kinga, D., Adam, J.B.: A method for stochastic optimization . In: International Conference on Learning Representations (ICLR). vol. 5 (2015)

work page 2015

Show all 18 references

[9]

ACM Transactions on Graphics (TOG) 33(6), 220 (2014)

Loper, M., Mahmood, N., Black, M.J.: Mosh: Motion and shape capture from sparse markers. ACM Transactions on Graphics (TOG) 33(6), 220 (2014)

2014
[10]

In: 29th British Machine Vision Conference (Sep 2018)

Ranjan, A., Romero, J., Black, M.J.: Learning human optical ﬂow. In: 29th British Machine Vision Conference (Sep 2018)

2018
[11]

vol- ume 1

Robinette, K.M., Blackwell, S., Daanen, H., Boehmer, M., F leming, S.: Civilian american and european surface anthropometry resource (caesar), ﬁnal report. vol- ume 1. summary. Tech. rep., DTIC Document (2002)

2002
[12]

IEEE transactions on pattern anal ysis and machine intelligence (2019)

Rogez, G., Weinzaepfel, P., Schmid, C.: Lcr-net++: Multi- person 2d and 3d pose detection in natural images. IEEE transactions on pattern anal ysis and machine intelligence (2019)

2019
[13]

ACM Transactions on Graphics, (Proc

Romero, J., Tzionas, D., Black, M.J.: Embodied hands: Mode ling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) 36(6) (Nov 2017)

2017
[14]

International journal of computer vision 87(1-2), 4 (2010)

Sigal, L., Balan, A.O., Black, M.J.: Humaneva: Synchroni zed video and motion capture dataset and baseline algorithm for evaluation of articu lated human motion. International journal of computer vision 87(1-2), 4 (2010)

2010
[15]

arXiv preprint arXiv:1409.1556 (2014)

Simonyan, K., Zisserman, A.: Very deep convolutional netw orks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)

2014 arXiv
[16]

In: Euro graphics

Teschner, M., Kimmerle, S., Heidelberger, B., Zachmann, G., Raghupathi, L., Fuhrmann, A., Cani, M.P., Faure, F., Magnenat-Thalmann, N., S trasser, W., Volino, P.: Collision detection for deformable objects. In: Euro graphics. pp. 119– 139 (2004)

2004
[17]

International Journal of Computer Vision (IJCV) 118(2), 172–193 (Jun 2016), https://doi.org/10.1007/s11263-016-0895-4

Tzionas, D., Ballan, L., Srikantha, A., Aponte, P., Pollefe ys, M., Gall, J.: Cap- turing hands in action using discriminative salient points an d physics simulation. International Journal of Computer Vision (IJCV) 118(2), 172–193 (Jun 2016), https://doi.org/10.1007/s11263-016-0895-4

2016 doi
[18]

In: CVPR (2017)

Varol, G., Romero, J., Martin, X., Mahmood, N., Black, M.J., Laptev, I., Schmid, C.: Learning from synthetic humans. In: CVPR (2017)

2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

http://mocap.cs.cmu.e du

Carnegie-mellon mocap database. http://mocap.cs.cmu.e du

work page

[2] [2]

https://github.com/matterport/Mask RCNN (2017)

Abdulla, W.: Mask r-cnn for object detection and instance seg mentation on keras and tensorﬂow. https://github.com/matterport/Mask RCNN (2017)

work page 2017

[3] [3]

In: Proceedings of the IEEE Conference on Compute r Vision and Pattern Recognition

Akhter, I., Black, M.J.: Pose-conditioned joint angle limi ts for 3d human pose reconstruction. In: Proceedings of the IEEE Conference on Compute r Vision and Pattern Recognition. pp. 1446–1455 (2015)

work page 2015

[4] [4]

In: 2017 IEEE Conference on Comput er Vision and Pattern Recognition (CVPR)

Cao, Z., Simon, T., Wei, S.E., Sheikh, Y.: Realtime multi- person 2d pose estima- tion using part aﬃnity ﬁelds. In: 2017 IEEE Conference on Comput er Vision and Pattern Recognition (CVPR). pp. 1302–1310. IEEE (2017)

work page 2017

[5] [5]

In: Pro- ceedings of the European Conference on Computer Vision (ECCV)

Fabbri, M., Lanzi, F., Calderara, S., Palazzi, A., Vezzani, R ., Cucchiara, R.: Learn- ing to detect and track visible and occluded body joints in a virt ual world. In: Pro- ceedings of the European Conference on Computer Vision (ECCV). pp. 430–446 (2018)

work page 2018

[6] [6]

Archives of the Game Developers Conference (Mar 2003), http://www.research.scea.com/gdc2003/spherical-harmonic-lighting.pdf

Green, R.: Spherical Harmonic Lighting: The Gritty Details. Archives of the Game Developers Conference (Mar 2003), http://www.research.scea.com/gdc2003/spherical-harmonic-lighting.pdf

work page 2003

[7] [7]

In: Computer Vision (ICCV), 2017 IEEE International Conference on

He, K., Gkioxari, G., Doll´ ar, P., Girshick, R.: Mask r-cnn. In: Computer Vision (ICCV), 2017 IEEE International Conference on. pp. 2980–2988. I EEE (2017)

work page 2017

[8] [8]

In: International Conference on Learning Representations (ICLR)

Kinga, D., Adam, J.B.: A method for stochastic optimization . In: International Conference on Learning Representations (ICLR). vol. 5 (2015)

work page 2015