Pith. sign in

REVIEW 3 major objections 7 minor 44 references

Question-guided mask selection lifts answer grounding to 65.3 IoU

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-07-31 23:31 UTC pith:NHB7VJID

load-bearing objection A plausible answer-grounding architecture whose headline dynamic-mask benefit is contradicted by its own fixed-resolution baseline, and whose SOTA numbers are unverifiable without code. the 3 major comments →

arxiv 2607.23921 v1 pith:NHB7VJID submitted 2026-07-27 cs.CV

DDVT: Dynamic Dual-level Vision Transformer Fusion Network for Answer Grounding in Visual Question Answering

Yue Zhang , Xiangyu Li , Wanshu Fan , Xin Yang , Dongsheng Zhou This is my paper
classification cs.CV
keywords answer groundingvisual question answeringdynamic networksvision transformermulti-scale fusioncross-modal attentionmask selectionGumbel-Softmax
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that answer grounding in visual question answering improves substantially when a network dynamically selects the resolution of the grounding mask based on the question, and when pixel-level and region-level visual features are fused under language guidance. It claims that this dual-level design, called DDVT, achieves 65.3 average IoU on VizWizGround and 43.47 on VQS, outperforming all cited baselines, including a 12-point gain over the nearest prior method on VizWizGround. The authors argue that the gain comes from two specific modules: a question-guided dynamic regional-level module (QGDR) that chooses among four mask resolutions via Gumbel-Softmax, and a cross-modal multi-scale aggregation module (CMA) that aligns and merges region-level features with pixel-level word-attended features. Ablations attribute 4.17 IoU to QGDR and 3.57 IoU to CMA on the validation set, supporting the claim that both components matter. A sympathetic reader would care because answer grounding is a step toward interpretable VQA, where the model can show which image region justifies its answer.

Core claim

The paper's central claim is that answer grounding accuracy improves by jointly predicting the answer and a grounding mask through a dual-level vision transformer fusion. The architecture starts with a vision transformer backbone and a language encoder, applies pixel-word attention for spatial alignment, then uses QGDR to locate region-level objects and stochastically select one of four mask resolutions (14×14 to 112×112) conditioned on the question. The selected region features are aligned and aggregated with pixel-level features in CMA using deformable convolution and conditionally parameterized convolution. Trained with mask, edge, budget, and text-answer losses, the model reports 65.3 Io

What carries the argument

The load-bearing components are QGDR and CMA. QGDR is a lightweight question-guided classifier that takes ROI-aligned region features, applies cross-modal attention with the question, channel attention, and outputs a probability vector over four mask resolutions; Gumbel-Softmax converts this into a differentiable one-hot selection. CMA is a cross-modal multi-scale aggregation module: it up-samples the QGDR region feature, concatenates it with the pixel-level word-attended feature, predicts an offset field, uses deformable convolution to align the region feature to the pixel feature, then fuses them with 1×1 convolution and conditionally parameterized convolution. Together they create the dua

Load-bearing premise

The 12-point lead over prior work is only as strong as the comparability of the cited baseline scores — if any baseline used a different test split, preprocessing, or IoU convention, the margin could shrink or vanish; separately, the paper's own data show a fixed maximum-resolution mask outperforming its dynamic selector, so the dynamic module's contribution is not established by its own experiments.

What would settle it

Run DDVT and the top baselines (DDTN, UNIFIED, or other published methods) on the same VizWizGround test split with identical preprocessing and IoU evaluation; if DDVT's margin over DDTN drops well below the reported 12 IoU points, the headline claim fails. A second, independent check: fix the mask resolution to 112×112 while keeping everything else identical — if that static variant matches or exceeds the full QGDR model on the validation set, as Table 3 already hints (74.68 vs 73.61), then the dynamic resolution mechanism is not the source of the reported improvement.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the reported numbers hold under a shared evaluation protocol, DDVT would set a new state of the art on VizWizGround for answer grounding with an IoU of 65.3, surpassing even a large unified model without large-scale pre-training.
  • On the VQS dataset, the method reaches 43.47 IoU, a clear margin over four cited baselines, suggesting the dual-level fusion transfers beyond the VizWizGround domain.
  • The ablations show that removing QGDR costs 4.17 IoU and replacing CMA with plain concatenation costs 3.57 IoU on the validation set, so both modules are needed for the claimed performance.
  • Varying the fixed mask size shows that larger masks improve IoU (from 70.82 at 14×14 to 74.68 at 112×112) at higher FLOPs; the dynamic mechanism is intended to approach large-mask quality at a lower average cost.
  • Qualitative results indicate the model produces sharper object boundaries and better captures global object information than the prior DDTN method, especially for complex contours such as plants.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural testable extension is to apply the QGDR resolution-selection idea to referring image segmentation or open-vocabulary grounding, where object scale varies widely and a per-instance resolution choice might give a similar efficiency-accuracy trade-off.
  • The paper's own Table 3 shows that a static 112×112 mask (74.68 IoU) outperforms the full dynamic model (73.61 IoU) on the validation set; this suggests the reported QGDR gain may come more from the auxiliary edge loss or overall network design than from the dynamic resolution choice, which a follow-up experiment could isolate.
  • Budget-constrained, input-dependent mask selection could be adapted to edge-device VQA, where per-image FLOPs must stay under a hard cap while preserving grounding quality.
  • The acknowledged gap between localization accuracy and textual answer accuracy points to a concrete follow-up: adding a less grounding-specialized text decoder or multi-task balancing could close the discrepancy without sacrificing mask IoU.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes DDVT, a dual-level vision-transformer fusion network for answer grounding in VQA. DDVT combines Swin Transformer visual features and BERT question features through a question-guided dynamic regional-level module (QGDR) that performs ROI-Align-based region selection with Gumbel-Softmax routing over multiple mask resolutions, and a cross-modal multi-scale aggregation module (CMA) that fuses pixel-level and region-level features. The authors report state-of-the-art IoU on VizWizGround (65.3) and VQS (43.47), with ablations attributing large gains to QGDR and CMA. The main technical contributions are the dynamic mask-resolution selection and the cross-modal fusion mechanism.

Significance. If the reported results are reproducible, the proposed architecture would represent a strong result on two answer-grounding benchmarks, and the dual-level fusion with language-guided dynamic selection would be a useful design direction. The paper uses standard, well-motivated components (Swin, BERT, LAVT-style PWAM, ROI Align, Gumbel-Softmax) and presents systematic ablations on the VizWizGround validation set. However, the significance is tempered by the absence of code or error bars, and by an internal inconsistency in the dynamic-selection evaluation that weakens the claimed contribution of the QGDR module.

major comments (3)
  1. [§4.5, Table 3 and §4.4, Table 2] Table 3 shows that a static 112×112 mask achieves 74.68 IoU on the VizWizGround val set, while the full dynamic model in Table 2 row (d) achieves 73.61 IoU. This is a same-set, same-metric comparison that directly contradicts the narrative that dynamic mask-resolution selection improves localization accuracy. The paper's own text in §4.5 frames the dynamic method as achieving only 'comparable performance at a lower cost,' yet Sections 1 and 3.1 present dynamic selection as a mechanism for improving grounding performance. Moreover, the 4.17 IoU difference between rows (a) and (d) in Table 2 conflates the dynamic routing with the ROI-Align features and cross-modal attention that are also part of QGDR. To support the claimed benefit of dynamic selection, the authors should provide an ablation that isolates the routing mechanism (e.g., full model with fixed 112×112 output vs. full model with
  2. [§4.3, Table 1] The headline claim of a 12% improvement over DDTN and state-of-the-art performance on VizWizGround and VQS is based on a single run, with no error bars or statistical significance tests. The baseline numbers are taken from other papers, and there is no shared evaluation harness, code release, or per-method description of test split, preprocessing, or IoU convention. If any baseline used a different protocol, the reported margin could be substantially reduced or vanish. The authors should report mean±std over multiple seeds, clearly describe the evaluation protocol, and ideally release code so that baselines can be re-run under identical conditions.
  3. [§4.2 and §4.3] The paper states in §4.3 that DDVT achieves state-of-the-art 'even without extensive pre-training,' yet §4.2 specifies that the Swin Transformer is initialized with ImageNet-22K classification weights and BERT with official pre-trained weights. The 'Pre-trained' column in Table 1 is not defined, and its symbols are inconsistent with this description. The authors should clarify what they mean by 'extensive pre-training' and define the column, otherwise the claim and the table are misleading.
minor comments (7)
  1. [§4.3, Fig. 2] Text in §4.3 refers to 'DDTN [29]' while Table 1 and Fig. 2 use [43]; reference 29 is a different work. Unify the citation.
  2. [Table 2 caption] The symbols '%' and '"' are used but not defined in the caption. The reader cannot tell which rows disable a module without guessing.
  3. [Eq. (4)] The notation is confusing: 'where hat m^k_i denotes the k-th mask prediction ground truth answer' is unclear. Presumably hat m^k_i is the predicted mask and m_i is the ground truth. Please fix the description.
  4. [§3.1] The sentence 'the obtained is performed with a global pooling operation' is incomplete. It should state which tensor is pooled.
  5. [§4.1] 'The range of the Q value is from 0 to 1' should say 'IoU value.'
  6. [§4.4] Typo: 'benefits benefits' in the ablation paragraph.
  7. [Table 3] The full dynamic model is not included in Table 3, making the comparison between static and dynamic resolution incomplete. Add the full-model row for reference.

Circularity Check

0 steps flagged

No circular derivation found: DDVT's IoU results are external measured comparisons; the flagged Table 3 inconsistency is an empirical problem, not a circular reduction.

full rationale

Walking the claimed derivation chain shows no load-bearing step that reduces to its own inputs. The headline result (Sec. 4.3, Table 1) is a measured IoU on held-out VizWizGround test and VQS val sets, compared with numbers cited from prior papers; it is not obtained by fitting a parameter to the target result, so there is no fitted-input-called-prediction pattern. QGDR's Gumbel-Softmax selection (Eq. 2), mask/edge losses (Eqs. 4-5), and budget constraint (Eq. 6) are standard supervised training components; the budget-constraint idea is explicitly cited to DynaMask [21], an externally published and separately evaluated method, so even if overlapping authorship exists it is not a self-referential load-bearing premise. CMA is assembled from cited external components (LAVT/PWAM [37], Deformable Conv [6], CondConv [36]). No uniqueness theorem is imported from the authors' prior work, and no known empirical pattern is merely renamed as a new result. The paper itself flags a relevant limitation: VizWizGround and VQS 'do not offer text answer verification' (Sec. 4.3), so Table 1 is an IoU-only comparison. Also, Table 3 shows a static 112x112 mask (74.68 IoU) beating the full dynamic model (73.61), which undercuts the claimed QGDR dynamic-selection benefit. These are baseline-comparability and internal-consistency concerns, not demonstrations that a prediction equals its construction inputs; they do not raise the circularity score.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claims rest on the transfer quality of pretrained Swin/BERT features, the reliability of VizWizGround/VQS mask annotations and their online evaluation, and most load-bearing, on the comparability of the cited baseline numbers in Table 1, which the paper does not regenerate. No new physical or formal entities are introduced; QGDR and CMA are module designs, not invented entities. Six hand-set or unreported design hyperparameters (loss weights, mask-resolution set, Gumbel temperature, budget target, ROI count) gate the reported accuracy and efficiency numbers.

free parameters (6)
  • lambda_1 (edge-loss weight) = 0.1
    Trade-off hyperparameter in Eq. (3); hand-set, no sensitivity analysis reported (Section 4.2).
  • lambda_2 (budget-loss weight) = 0.4
    Trade-off hyperparameter in Eq. (3); hand-set, no sensitivity analysis reported (Section 4.2).
  • candidate mask resolutions = {14, 28, 56, 112}
    The QGDR candidate resolution set is chosen by hand; Table 3 shows larger masks monotonically improve IoU, so this choice partly determines the reported accuracy (Section 4.5).
  • Gumbel-Softmax temperature tau = not reported
    Eq. (2) requires tau for the one-hot approximation; the value is never given, so the dynamic selection mechanism is underspecified.
  • budget target Ct = not reported
    Eq. (6) penalizes expected computational cost above Ct; the target value is never stated, so the claimed efficiency trade-off cannot be reproduced.
  • ROI proposal count = 50 proposals in, 30 kept
    Fig. 5 caption: QGDR scores 50 ROI-aligned boxes and keeps 30; the proposal budget is a hand-chosen design choice affecting localization quality.
axioms (5)
  • domain assumption Swin Transformer visual features and BERT question embeddings provide adequate paired cross-modal representations for grounding.
    Invoked throughout Section 3; the entire fusion design assumes these pretrained encoders carry the semantics needed for mask localization.
  • domain assumption ImageNet-22K and BERT pretraining transfer to the VizWizGround and VQS grounding tasks.
    Section 4.2 initializes Swin with ImageNet-22K weights and BERT with official weights; the reported gains implicitly assume this transfer.
  • domain assumption VizWizGround and VQS mask annotations and the IoU/mAP evaluation protocol faithfully measure answer-grounding quality.
    Section 4.1 defines the metrics; the claim of superiority inherits the quality and comparability of these benchmark annotations.
  • domain assumption The cited baseline scores in Table 1 (LXMERT, Mac-Caps, UNIFIED, DDTN, SDCAM, etc.) were produced under evaluation conditions comparable to the authors' own runs.
    Section 4.3 relies entirely on cross-paper numbers for the 12% claim; no protocol, split, or preprocessing description is given for those baselines, and no code is released to check comparability.
  • standard math Gumbel-Softmax (Eq. 2) provides usable gradients for discrete mask-resolution selection.
    Section 3.1 relies on the standard reparameterization; this is a routine mathematical assumption, not specific to the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 9511 in / 20429 out tokens · 190405 ms · 2026-07-31T23:31:53.328822+00:00 · methodology

0 comments
read the original abstract

Answer grounding in visual question answering aims to locate the region from a given natural language question associated with the visual content of an image, which has garnered significant attention due to its practical applications. In this paper, we introduce the Dynamic Dual-level Vision Transformer Fusion Network (DDVT) for answer grounding in visual question answering. Specifically, we propose a question-guided dynamic regional-level module (QGDR) that combines complementary image context through ROI Align and text content, enabling precise localization of text-related visual content. Moreover, we present a cross-modal multi-scale aggregation module (CMA) that enhances feature fusion between pixel-level and region-level features, facilitating the effective localization of visual content associated with grounded answers. Furthermore, we fuse the located visual content with text features to locate the region and provide answers to questions posed about the image. Experimental results demonstrate that our DDVT outperforms state-of-the-art methods on several widely-used benchmarks.

Figures

Figures reproduced from arXiv: 2607.23921 by Dongsheng Zhou, Wanshu Fan, Xiangyu Li, Xin Yang, Yue Zhang.

Figure 1
Figure 1. Figure 1: The overall architecture of Dynamic Dual-level Vision Transformer Fusion Network (DDVT). we leverage the Swin Transformer to extract visual features and the BERT model to obtain language question features. These are then fused using our pro￾posed Question-Guided Dynamic Regional-Level (QGDR) module, enabling dynamic object region identification and mask selection based on questions. Additionally, we introd… view at source ↗
Figure 2
Figure 2. Figure 2: Predicted masks generated by DDTN [43] and our DDVT on the Vizwiz￾Ground [4] val set. The range of the Q value is from 0 to 1, with higher values indicating better performance. For the validation set, we also evaluate using the general metrics for detection and localization tasks: mAP@IoU based on the COCO evaluation protocol, and the average AP values with IoU thresholds ranging from 0.5 to 0.95 with a st… view at source ↗
Figure 3
Figure 3. Figure 3: Visualized predictions from the VizWizground [4] val set. Baseline is LAVT [37] [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualized predictions and feature maps on an example from the VizWizgroun validation set. (a) shows from left to right, the left-most column illustrates the input image and the truth answer ground mask. We visualize the predicted mask and the feature maps used for final classification (i.e., Y4,Y3, Y2, and Y1) from left to right. (b) shows the output of attention to questions in word pixel attention, with… view at source ↗
Figure 5
Figure 5. Figure 5: (a) demonstrates the output bounding boxes of 50 after being processed through ROI Align, as well as the output bounding boxes of 30 after being scored by QGDR. (b) displays the cross-attention maps generated by QGDR. Each index from [1-30] corresponds to each object in the subgraph (a). 12 [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: The visualization in Fig. 5 (a) displays 50 boxes detected by the Roi Align [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 2 linked inside Pith

  1. [1]

    In: Proc

    Anderson, P., He, X., Buehler, C.: Bottom-up and top-down attention for image captioning and visual question answering. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 6077–6086 (2018)

    2022

  2. [2]

    In: Proc

    Ben Abacha, A., Hasan, S.A., Datla, V.V., Demner-Fushman, D., Müller, H.: Vqa- med: Overview of the medical visual question answering task at imageclef 2019. In: Proc. of Conf. and Labs of the Evaluation Forum. 9-12 September 2019 (2019) 13

    2019

  3. [3]

    In: Proc

    Bolukbasi, T., Wang, J., Dekel, O., Saligrama, V.: Adaptive neural networks for efficient inference. In: Proc. of the 2022 INt. Conf. on Mach. Learn. pp. 527–536 (2017)

    2022

  4. [4]

    In: Proc

    Chen, C., Anjum, S., Gurari, D.: Grounding answers for visual questions asked by visually impaired people. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 19098–19107 (2022)

    2022

  5. [5]

    In: Proc

    Cheng, T., Wang, X., Huang, L., Liu, W.: Boundary-preserving mask r-cnn. In: Proc. of the 2022 Eur. Conf. on Comput. Vis. pp. 660–676 (2020)

    2022

  6. [6]

    In: Proc

    Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable con- volutional networks. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 764–773 (2017)

    2022

  7. [7]

    In: Proc

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 248–255 (2009)

    2022

  8. [8]

    In: Proc

    Gan, C., Li, Y., Li, H., Sun, C., Gong, B.: Vqs: Linking segmentations to ques- tions and answers for supervised attention in vqa and question-focused semantic segmentation. In: Proc. of the 2022 IEEE/CVF Int. Conf. on Comput. Vis. pp. 1811–1820 (2017)

    2022

  9. [9]

    Gan, C., Li, Y., Li, H., Sun, C., Gong, B.: Vqs: Linking segmentations to ques- tions and answers for supervised attention in vqa and question-focused semantic segmentation. pp. 1829–1838 (2017)

    2017

  10. [10]

    In: Proc

    Gao, P., Jiang, Z., You, H., Lu, P., Hoi, S.C.H., Wang, X., Li, H.: Dynamic fusion with intra- and inter-modality attention flow for visual question answering. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 6632–6641 (2019)

    2022

  11. [11]

    IEEE Trans

    Han, Y., Huang, G., Song, S., Yang, L., Wang, H., Wang, Y.: Dynamic neural networks: A survey. IEEE Trans. Pattern Anal. Mach. Intell.44, 7436–7456 (2021)

    2021

  12. [12]

    In: Proc

    Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 7132–7141 (2018)

    2022

  13. [13]

    In: Proc

    Hu, R., Rohrbach, A., Darrell, T., Saenko, K.: Language-conditioned graph net- works for relational reasoning. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 10294–10303 (2019)

    2022

  14. [14]

    In: Proc

    Huang, G., Chen, D., Li, T., Wu, F., van der Maaten, L., Weinberger, K.Q.: Multi- scale dense networks for resource efficient image classification. In: Proc. of the 2022 Int. Conf. on Learn. Represent. (2018)

    2022

  15. [15]

    In: Proc

    Hudson, D.A., Manning, C.D.: Gqa: A new dataset for real-world visual reasoning and compositional question answering. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 6700–6709 (2019)

    2022

  16. [16]

    In: Proc

    Jang, E., Gu, S., Poole, B.: Categorical reparameterization with gumbel-softmax. In: Proc. of the 2022 Int. Conf. on Learn. Represent. (2017)

    2022

  17. [17]

    Jiang, Y., Natarajan, V., Chen, X., Rohrbach, M., Batra, D., Parikh, D.: Pythia v0.1:thewinningentrytotheVQAchallenge2018.CoRRabs/1807.09956(2018)

    Pith/arXiv arXiv 2018

  18. [18]

    In: Proc

    Khan, A.U., Kuehne, H., Duarte, K., Gan, C., da Vitoria Lobo, N., Shah, M.: Found a reason for me? weakly-supervised grounded visual question answering using capsules. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 8465–8474 (2021)

    2022

  19. [19]

    In: EMNLP

    Khan, A.U., Mazaheri, A., da Vitoria Lobo, N., Shah, M.: MMFT-BERT: multi- modal fusion transformer with BERT encodings for visual question answering. In: EMNLP. pp. 4648–4660 (2020)

    2020

  20. [20]

    IEEE Signal Process

    Lan, H., Zhang, P.: Learning and integrating multi-level matching features for image-text retrieval. IEEE Signal Process. Lett.29, 374–378 (2022) 14

    2022

  21. [21]

    In: Proc

    Li, R., He, C., Li, S., Zhang, Y., Zhang, L.: Dynamask: Dynamic mask selection for instance segmentation. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 11279–11288 (2023)

    2022

  22. [22]

    IEEE Signal Process

    Li, W., Ma, Z., Shi, J., Fan, X.: The style transformer with common knowledge optimization for image-text retrieval. IEEE Signal Process. Lett.30, 1197–1201 (2023)

    2023

  23. [23]

    NIPS30(2017)

    Lin, J., Rao, Y., Lu, J., Zhou, J.: Runtime neural pruning. NIPS30(2017)

    2017

  24. [24]

    IEEE Signal Process

    Liu, Y., Liu, H., Wang, H., Liu, M.: Regularizing visual semantic embedding with contrastive learning for image-text matching. IEEE Signal Process. Lett.29, 1332– 1336 (2022)

    2022

  25. [25]

    In: Proc

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proc. of the 2022 IEEE/CVF Int. Conf. on Comput. Vis. pp. 10012–10022 (2021)

    2022

  26. [26]

    In: Proc

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: Proc. of the 2022 Int. Conf. on Learn. Represent. (2019)

    2022

  27. [27]

    In: Proc

    Lu, J., Clark, C., Zellers, R., Mottaghi, R., Kembhavi, A.: UNIFIED-IO: A unified model for vision, language, and multi-modal tasks. In: Proc. of the 2022 Int. Conf. on Learn. Represent. (2023)

    2022

  28. [28]

    In: Proc

    Mullapudi, R.T., Mark, W.R., Shazeer, N., Fatahalian, K.: Hydranets: Specialized dynamic architectures for efficient inference. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 8080–8089 (2018)

    2022

  29. [29]

    CoRR abs/2207.05703(2022)

    Pan, J., Chen, G., Liu, Y., Wang, J., Bian, C., Zhu, P., Zhang, Z.: Tell me the evidence? dual visual-linguistic interaction for answer grounding. CoRR abs/2207.05703(2022)

    Pith/arXiv arXiv 2022

  30. [30]

    Annual Conference on Neural Information Pro- cessing Systems32(2019)

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Annual Conference on Neural Information Pro- cessing Systems32(2019)

    2019

  31. [31]

    Sensors20, 2281 (2020)

    Qiu, Y., Satoh, Y., Suzuki, R., Iwata, K., Kataoka, H.: Multi-view visual question answering with active viewpoint selection. Sensors20, 2281 (2020)

    2020

  32. [32]

    Tan, H., Bansal, M.: LXMERT: learning cross-modality encoder representations from transformers. In: Pro. of the 2019 Conf. on Empirical Methods in Natural Language Processing and the 9th International Joint Conf. on Natural Language Processin. pp. 5099–5110 (2019)

    2019

  33. [33]

    In: Proc

    Tian, W., Li, H., Zhao, Z.Q.: Dual capsule attention mask network with mutual learning for visual question answering. In: Proc. of the 29th International Conf. on Computational Linguistics. pp. 5678–5688 (2022)

    2022

  34. [34]

    In: Proc

    Wang, X., Yu, F., Dou, Z.Y., Darrell, T., Gonzalez, J.E.: Skipnet: Learning dy- namic routing in convolutional networks. In: Proc. of the 2022 Eur. Conf. on Com- put. Vis. pp. 409–424 (2018)

    2022

  35. [35]

    In: EMNLP

    Wolf, T., Debut, L., Sanh, V., et al., J.C.: Transformers: State-of-the-art natural language processing. In: EMNLP. pp. 38–45 (2020)

    2020

  36. [36]

    Annual Conference on Neural Information Processing Systems32(2019)

    Yang, B., Bender, G., Le, Q.V., Ngiam, J.: Condconv: Conditionally parameter- ized convolutions for efficient inference. Annual Conference on Neural Information Processing Systems32(2019)

    2019

  37. [37]

    In: Proc

    Yang, Z., Wang, J., Tang, Y., Chen, K., Zhao, H., Torr, P.H.S.: LAVT: language- aware vision transformer for referring image segmentation. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 18134–18144 (2022)

    2022

  38. [38]

    Annual Conference on Neural Information Processing Systems 201932(2019) 15

    You, Z., Yan, K., Ye, J., Ma, M., Wang, P.: Gate decorator: Global filter pruning method for accelerating deep convolutional neural networks. Annual Conference on Neural Information Processing Systems 201932(2019) 15

    2019

  39. [39]

    In: Proc

    Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 6281–6290 (2019)

    2022

  40. [40]

    In: Proc

    Yu, Z., Yu, J., Cui, Y., Tao, D., Tian, Q.: Deep modular co-attention networks for visual question answering. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 6274–6283 (2019)

    2022

  41. [41]

    In: IJCAI Int

    Zhang, S., Qu, L., You, S., Yang, Z., Zhang, J.: Automatic generation of grounded visual questions. In: IJCAI Int. Joint Conf. Artif. Intell. Int. Joint Conf. Artif. Intell. pp. 4235–4243 (2017)

    2017

  42. [42]

    In: Proc

    Zhou, Y., Ren, T., Zhu, C., Sun, X., Liu, J.: Trar: Routing the attention spans in transformer for visual question answering. In: Proc. of the 2022 IEEE Int. Conf. on Comput. Vis. and Pattern Recognit. pp. 2074–2084 (2021)

    2022

  43. [43]

    Pattern Recognit

    Zhu,L.,Peng,L.,Zhou,W.,Yang,J.:Dual-decodertransformernetworkforanswer grounding in visual question answering. Pattern Recognit. Lett.171, 53–60 (2023)

    2023

  44. [44]

    Annual Conference on Neural Information Processing Systems34, 27319–27330 (2021) 16

    Zhu, M., Han, K., Wu, E., Zhang, Q., Nie, Y., Lan, Z., Wang, Y.: Dynamic reso- lution network. Annual Conference on Neural Information Processing Systems34, 27319–27330 (2021) 16

    2021