Deep Instance-Level Hard Negative Mining Model for Histopathology Images

Arnold Wiliem; Brian C. Lovell; Kun Zhao; Lin Wu; Meng Li; Teng Zhang

arxiv: 1906.09681 · v3 · pith:RUAE3KWKnew · submitted 2019-06-24 · 💻 cs.CV · cs.AI

Deep Instance-Level Hard Negative Mining Model for Histopathology Images

Meng Li , Lin Wu , Arnold Wiliem , Kun Zhao , Teng Zhang , Brian C. Lovell This is my paper

Pith reviewed 2026-05-25 18:05 UTC · model grok-4.3

classification 💻 cs.CV cs.AI

keywords multiple instance learninghistopathologyattention mechanismhard negative miningwhole slide imagescancer classificationdeep CNN

0 comments

The pith

A CNN for histopathology slides learns attention weights on patches to classify bags and generate hard-negative instances for better accuracy.

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

Histopathology whole-slide images are treated as bags of patches under multiple instance learning, where the model must predict a single label for the bag. The paper builds a deep CNN that performs this bag classification while using an attention mechanism to assign weights to individual patches, revealing which ones drive the decision. Adaptive weighting during training shifts focus toward hard samples in each bag, and the learned attention scores are then used to construct new bags consisting of hard negative instances. This combination is meant to raise classification performance while adding instance-level interpretability. Experiments on colon and breast cancer datasets are reported to reach state-of-the-art results.

Core claim

The framework embeds an attention layer inside a CNN for MIL-based WSI classification so that the network both predicts the bag label and produces per-instance attention weights. These weights support two further steps: adaptive re-weighting of instances inside each training bag to emphasize difficult samples, and creation of additional bags populated with hard negative instances drawn according to the attention scores. The resulting model is shown to deliver state-of-the-art accuracy on colon and breast cancer histopathology collections.

What carries the argument

Attention mechanism that transforms instances and produces weights used both to locate key patches and to construct hard-negative bags, together with adaptive instance weighting inside each training bag.

If this is right

The attention weights supply instance-level explanations for each bag prediction.
Adaptive weighting forces the optimizer to pay more attention to difficult patches inside every bag.
Hard-negative bag generation augments the training distribution with challenging counter-examples.
State-of-the-art bag classification accuracy is obtained on colon and breast cancer histopathology data.

Where Pith is reading between the lines

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

If the attention weights align with regions marked by pathologists, the model could reduce reliance on pixel-level supervision in other medical imaging tasks.
The same attention-plus-hard-negative pattern could be tested on non-medical MIL problems such as document or video classification.
Performance may degrade if the assumption that every bag contains a mixture of positive and negative instances is violated.

Load-bearing premise

The attention weights correctly surface the patches that determine the bag label.

What would settle it

On the same colon and breast datasets, replacing the attention-derived hard-negative bags and adaptive weights with random instance selection produces no drop in classification accuracy.

Figures

Figures reproduced from arXiv: 1906.09681 by Arnold Wiliem, Brian C. Lovell, Kun Zhao, Lin Wu, Meng Li, Teng Zhang.

**Figure 1.** Figure 1: The architecture of the end-to-end deep CNN model with adaptive attention mechanism. The input is the bag of instances (patches of each WSI), which are fed into a CNN model to produce the latent representation of each instance. Then the embeddings of instances go through a fully connected network with the attention weights generated. The learned weights are multiplied with the embeddings of instances in el… view at source ↗

**Figure 2.** Figure 2: The proposed novel hard negative mining process. The training images are first fed into the deep MIL model with balanced training to select instances to constitute the false positive bags. We learn attention weights for instances, which can be used to select the hard instances that fool the model to make the wrong prediction. Next, the hard negative instances are grouped to form the new hard negative bags … view at source ↗

**Figure 3.** Figure 3: Positive and hard negative examples: (a) Colon data: instances that include malignant regions. (b) Colon data: detected hard negative instances that mislead the model to predict a normal bag into a malignant result. (c) UCSB data: instances that include malignant regions. (d) UCSB data: detected hard negative instances that mislead the model to predict a normal bag into a malignant result. than the others … view at source ↗

read the original abstract

Histopathology image analysis can be considered as a Multiple instance learning (MIL) problem, where the whole slide histopathology image (WSI) is regarded as a bag of instances (i.e, patches) and the task is to predict a single class label to the WSI. However, in many real-life applications such as computational pathology, discovering the key instances that trigger the bag label is of great interest because it provides reasons for the decision made by the system. In this paper, we propose a deep convolutional neural network (CNN) model that addresses the primary task of a bag classification on a WSI and also learns to identify the response of each instance to provide interpretable results to the final prediction. We incorporate the attention mechanism into the proposed model to operate the transformation of instances and learn attention weights to allow us to find key patches. To perform a balanced training, we introduce adaptive weighing in each training bag to explicitly adjust the weight distribution in order to concentrate more on the contribution of hard samples. Based on the learned attention weights, we further develop a solution to boost the classification performance by generating the bags with hard negative instances. We conduct extensive experiments on colon and breast cancer histopathology data and show that our framework achieves state-of-the-art performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Incremental attention-MIL plus hard-negative mining for histopathology, but the SOTA claim has no numbers attached and the attention step lacks any validation.

read the letter

The paper puts together an attention-weighted MIL CNN, adaptive per-bag weighting, and attention-guided construction of hard-negative bags for whole-slide histopathology classification. The end-to-end pipeline is new for this exact use case, though each component has appeared before. It does a reasonable job of spelling out why instance-level explanations matter in computational pathology, where clinicians need to see which patches support the slide label. That motivation is clear and on target for the domain. The adaptive weighting step to balance hard samples inside each bag is a practical addition that could help training stability. The main weaknesses are straightforward. The abstract asserts state-of-the-art results on colon and breast cancer datasets yet shows zero numbers, baselines, error bars, or statistical tests, so the performance claim cannot be checked. The stress-test point about attention validation is on point: bag-level labels only mean the attention module has no direct supervision, so the weights used to pick hard-negative instances could be arbitrary. If they do not track actual diagnostic regions, the hard-negative step becomes ordinary resampling and the claimed mechanism does not explain any gains. No circular reasoning or invented quantities appear. The work is aimed at people already doing MIL on pathology slides who might want to try this particular combination. A reader in that niche could extract implementation details, but the paper does not introduce a result that stands on its own without the missing tables. It is worth sending to peer review so the experiments can be examined properly; the description is clear enough to referee even if revisions will be required to substantiate the claims.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes a deep CNN model for multiple instance learning (MIL) on whole-slide histopathology images (WSIs), treating each WSI as a bag of patches. It integrates an attention mechanism to learn instance weights for identifying key patches, applies adaptive per-bag weighting to emphasize hard samples, and generates hard-negative bags using the learned attention weights to boost classification. The authors report extensive experiments on colon and breast cancer datasets and claim state-of-the-art performance.

Significance. If the empirical results and the link between attention weights and performance gains hold, the work would offer a practical extension of attention-based MIL to histopathology with potential for improved interpretability alongside accuracy. The absence of circular reasoning in the supervised pipeline is a methodological strength.

major comments (2)

[Abstract] Abstract: the central claim of state-of-the-art performance on colon and breast cancer histopathology data is asserted without any quantitative numbers, baselines, statistical tests, or error bars supplied in the manuscript text; this directly prevents evaluation of the empirical contribution.
[Abstract] Abstract (paragraph describing the attention and hard-negative generation steps): the construction of hard-negative bags and the adaptive weighting both rely on the assumption that attention weights correctly surface diagnostically decisive instances, yet no instance-level supervision or independent validation (e.g., overlap with pathologist annotations) is provided; without such grounding the performance benefit cannot be attributed to the claimed mechanism.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point by point below, proposing revisions where they strengthen the work without misrepresenting our contributions.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of state-of-the-art performance on colon and breast cancer histopathology data is asserted without any quantitative numbers, baselines, statistical tests, or error bars supplied in the manuscript text; this directly prevents evaluation of the empirical contribution.

Authors: We agree that the abstract would be strengthened by including specific quantitative results. In the revised manuscript, we will update the abstract to report key accuracy metrics, baseline comparisons, and references to statistical significance from our experiments on the colon and breast cancer datasets. revision: yes
Referee: [Abstract] Abstract (paragraph describing the attention and hard-negative generation steps): the construction of hard-negative bags and the adaptive weighting both rely on the assumption that attention weights correctly surface diagnostically decisive instances, yet no instance-level supervision or independent validation (e.g., overlap with pathologist annotations) is provided; without such grounding the performance benefit cannot be attributed to the claimed mechanism.

Authors: Our approach operates under standard weakly supervised MIL settings using only bag-level labels, as instance-level annotations are typically unavailable in histopathology. The attention weights are optimized end-to-end for bag classification, and ablation experiments in the paper demonstrate performance gains specifically from the attention-driven hard-negative bag generation and adaptive weighting. We will revise the abstract and add a limitations discussion clarifying that interpretability claims are based on the learned weights without external pathologist validation, while the empirical results support the mechanism's utility. revision: partial

Circularity Check

0 steps flagged

No circularity; standard supervised MIL pipeline with attention and hard-negative sampling

full rationale

The paper describes an end-to-end CNN trained on bag-level labels only. Attention weights are learned as part of the classification objective; hard-negative bags are then constructed from those weights and used in the same training loop. No equation or claim reduces a derived quantity to a fitted parameter by definition, no self-citation chain is invoked as a uniqueness theorem, and no ansatz is smuggled in. All steps remain externally falsifiable via held-out bag classification accuracy on colon and breast datasets. This is the normal case of a supervised deep-learning architecture.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the standard MIL bag-label assumption and the unverified premise that learned attention weights identify diagnostically causal patches; no free parameters, invented entities, or non-standard axioms are stated in the abstract.

axioms (1)

domain assumption Whole-slide histopathology images can be treated as bags whose single label is determined by a small subset of patches (MIL assumption)
Explicitly stated in the first two sentences of the abstract.

pith-pipeline@v0.9.0 · 5763 in / 1240 out tokens · 35842 ms · 2026-05-25T18:05:03.205083+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We incorporate the attention mechanism into the proposed model to operate the transformation of instances and learn attention weights... Based on the learned attention weights, we further develop a solution to boost the classification performance by generating the bags with hard negative instances.
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

zi = sum wij gij, wij = exp(...) / sum... adaptive weighting... hard negative instances through attention weights: Hl = {wli | wli >= sigma_l + wl}

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 1 internal anchor

[1]

Deep learning-based assessment of tumor-associated stroma for diagnosing breast cancer in histopathology images

Babak Ehteshami Bejnordi, Jimmy Lin, Ben Glass, Maeve Mul looly, Gretchen L Gierach, Mark E Sherman, Nico Karssemeijer, Jeroen Van Der L aak, and Andrew H Beck. Deep learning-based assessment of tumor-associated stroma for diagnosing breast cancer in histopathology images. In ISBI, 2017

work page 2017
[2]

Multiple instance learning for hetero geneous images: Train- ing a cnn for histopathology

Heather D Couture, James Stephen Marron, Charles M Perou, Melissa A Troester, and Marc Niethammer. Multiple instance learning for hetero geneous images: Train- ing a cnn for histopathology. In MICCAI, 2018

work page 2018
[3]

Histograms of oriented gra dients for human detec- tion

Navneet Dalal and Bill Triggs. Histograms of oriented gra dients for human detec- tion. In CVPR, 2005

work page 2005
[4]

Avoiding false positiv e in multi-instance learning

Yanjun Han, Qing Tao, and Jue Wang. Avoiding false positiv e in multi-instance learning. In NIPS, 2010

work page 2010
[5]

Kurc, Yi Gao, James E

Le Hou, Dimitris Samaras, Tahsin M. Kurc, Yi Gao, James E. D avis, and Joel H. Saltz. Patch-based Convolutional Neural Network for Whole Slide Tissue Image Classiﬁcation. In CVPR, 2016

work page 2016
[6]

Attenti on-based deep multiple instance learning

Maximilian Ilse, Jakub Tomczak, and Max Welling. Attenti on-based deep multiple instance learning. In ICML, 2018

work page 2018
[7]

Hamprecht

Melih Kandemir, Chong Zhang, and Fred A. Hamprecht. Empow ering multiple instance histopathology cancer diagnosis by cell graphs. I n MICCAI 2014

work page 2014
[8]

Where-and-When to Look: Deep Siamese Attention Networks for Video-based Person Re-iden tiﬁcation

Lin Wu, Yang Wang, Junbin Gao and Xue Li. Where-and-When to Look: Deep Siamese Attention Networks for Video-based Person Re-iden tiﬁcation. In IEEE Transactions on Multimedia , 2019

work page 2019
[9]

Deep Attention-ba sed Spatially Recursive Networks for Fine-Grained Visual Recognition

Lin Wu, Yang Wang, Xue Li and Junbin Gao. Deep Attention-ba sed Spatially Recursive Networks for Fine-Grained Visual Recognition. I n IEEE Transactions on Cybernetics, 49(5):1791–1802, 2019

work page 2019
[10]

Classifyin g and segmenting microscopy images with deep multiple instance learning

Oren Z Kraus, Jimmy Lei Ba, and Brendan J Frey. Classifyin g and segmenting microscopy images with deep multiple instance learning. In Bioinformatics, 2016

work page 2016
[11]

Key instance d etection in multi- instance learning

Guoqing Liu, Jianxin Wu, and Zhi-Hua Zhou. Key instance d etection in multi- instance learning. In ACML, 2012

work page 2012
[12]

A threshold selection method from gray-l evel histograms

Nobuyuki Otsu. A threshold selection method from gray-l evel histograms. In SMCS, 1979

work page 1979
[13]

Explaining t he stars: Weighted multiple-instance learning for aspect-based sentiment an alysis

Nikolaos Pappas and Andrei Popescu-Belis. Explaining t he stars: Weighted multiple-instance learning for aspect-based sentiment an alysis. In EMNLP, 2014

work page 2014
[14]

Explicit doc ument modeling through weighted multiple-instance learning

Nikolaos Pappas and Andrei Popescu-Belis. Explicit doc ument modeling through weighted multiple-instance learning. In JAIR, 2017

work page 2017
[15]

Training region-based object detectors with online hard example mining

Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016

work page 2016
[16]

Sirinukunwattana, S

K. Sirinukunwattana, S. E. A. Raza, Y. Tsang, D. R. J. Snea d, I. A. Cree, and N. M. Rajpoot. Locality sensitive deep learning for detecti on and classiﬁcation of nuclei in routine colon cancer histology images. In T-MI, 2016

work page 2016
[17]

Y. Song, Q. Li, H. Huang, D. Feng, M. Chen, and W. Cai. Low di mensional representation of ﬁsher vectors for microscopy image class iﬁcation. In T-MI, 2017

work page 2017
[18]

M. Sun, T. X. Han, , and A. Khodayari-Rostamabad. Multipl e instance learning convolutional neural networks for object recognition. In ICPR, 2016

work page 2016
[19]

Deep Learning for Identifying Metastatic Breast Cancer

Dayong Wang, Aditya Khosla, Rishab Gargeya, Humayun Irs had, and Andrew H Beck. Deep learning for identifying metastatic breast canc er. arXiv preprint arXiv:1606.05718, 2016

work page internal anchor Pith review Pith/arXiv arXiv 2016
[20]

Yan. Xu, J. Zhu, E. Chang, and Z. Tu. Multiple clustered in stance learning for histopathology cancer image classiﬁcation, segmentation and clustering. In CVPR, June 2012

work page 2012
[21]

Xu, Tao Mo, Qiwei Feng, Peilin Zhong, Maode Lai, I Eri c, and Chao Chang

Yan. Xu, Tao Mo, Qiwei Feng, Peilin Zhong, Maode Lai, I Eri c, and Chao Chang. Deep learning of feature representation with multiple inst ance learning for medical image analysis. In ICASSP, 2014

work page 2014
[22]

Deep adaptive fe ature embedding with local sample distributions for person re-identiﬁcati on

Lin Wu, Yang Wang, Junbin Gao, and Xue Li. Deep adaptive fe ature embedding with local sample distributions for person re-identiﬁcati on. Pattern Recognition, 73:275–288, 2018

work page 2018
[23]

Deep attention- based spatially recursive networks for ﬁne-grained visual recognition

Lin Wu, Yang Wang, Xue Li, and Junbin Gao. Deep attention- based spatially recursive networks for ﬁne-grained visual recognition. IEEE Transactions on Cy- bernetics, 49(5):1791–1802, 2019

work page 2019
[24]

Cycle-consistent deep g enerative hashing for cross-modal retrieval

Lin Wu, Yang Wang, and Ling Shao. Cycle-consistent deep g enerative hashing for cross-modal retrieval. IEEE Transactions on Image Processing , 28(4):1602–1612, 2019

work page 2019

[1] [1]

Deep learning-based assessment of tumor-associated stroma for diagnosing breast cancer in histopathology images

Babak Ehteshami Bejnordi, Jimmy Lin, Ben Glass, Maeve Mul looly, Gretchen L Gierach, Mark E Sherman, Nico Karssemeijer, Jeroen Van Der L aak, and Andrew H Beck. Deep learning-based assessment of tumor-associated stroma for diagnosing breast cancer in histopathology images. In ISBI, 2017

work page 2017

[2] [2]

Multiple instance learning for hetero geneous images: Train- ing a cnn for histopathology

Heather D Couture, James Stephen Marron, Charles M Perou, Melissa A Troester, and Marc Niethammer. Multiple instance learning for hetero geneous images: Train- ing a cnn for histopathology. In MICCAI, 2018

work page 2018

[3] [3]

Histograms of oriented gra dients for human detec- tion

Navneet Dalal and Bill Triggs. Histograms of oriented gra dients for human detec- tion. In CVPR, 2005

work page 2005

[4] [4]

Avoiding false positiv e in multi-instance learning

Yanjun Han, Qing Tao, and Jue Wang. Avoiding false positiv e in multi-instance learning. In NIPS, 2010

work page 2010

[5] [5]

Kurc, Yi Gao, James E

Le Hou, Dimitris Samaras, Tahsin M. Kurc, Yi Gao, James E. D avis, and Joel H. Saltz. Patch-based Convolutional Neural Network for Whole Slide Tissue Image Classiﬁcation. In CVPR, 2016

work page 2016

[6] [6]

Attenti on-based deep multiple instance learning

Maximilian Ilse, Jakub Tomczak, and Max Welling. Attenti on-based deep multiple instance learning. In ICML, 2018

work page 2018

[7] [7]

Hamprecht

Melih Kandemir, Chong Zhang, and Fred A. Hamprecht. Empow ering multiple instance histopathology cancer diagnosis by cell graphs. I n MICCAI 2014

work page 2014

[8] [8]

Where-and-When to Look: Deep Siamese Attention Networks for Video-based Person Re-iden tiﬁcation

Lin Wu, Yang Wang, Junbin Gao and Xue Li. Where-and-When to Look: Deep Siamese Attention Networks for Video-based Person Re-iden tiﬁcation. In IEEE Transactions on Multimedia , 2019

work page 2019

[9] [9]

Deep Attention-ba sed Spatially Recursive Networks for Fine-Grained Visual Recognition

Lin Wu, Yang Wang, Xue Li and Junbin Gao. Deep Attention-ba sed Spatially Recursive Networks for Fine-Grained Visual Recognition. I n IEEE Transactions on Cybernetics, 49(5):1791–1802, 2019

work page 2019

[10] [10]

Classifyin g and segmenting microscopy images with deep multiple instance learning

Oren Z Kraus, Jimmy Lei Ba, and Brendan J Frey. Classifyin g and segmenting microscopy images with deep multiple instance learning. In Bioinformatics, 2016

work page 2016

[11] [11]

Key instance d etection in multi- instance learning

Guoqing Liu, Jianxin Wu, and Zhi-Hua Zhou. Key instance d etection in multi- instance learning. In ACML, 2012

work page 2012

[12] [12]

A threshold selection method from gray-l evel histograms

Nobuyuki Otsu. A threshold selection method from gray-l evel histograms. In SMCS, 1979

work page 1979

[13] [13]

Explaining t he stars: Weighted multiple-instance learning for aspect-based sentiment an alysis

Nikolaos Pappas and Andrei Popescu-Belis. Explaining t he stars: Weighted multiple-instance learning for aspect-based sentiment an alysis. In EMNLP, 2014

work page 2014

[14] [14]

Explicit doc ument modeling through weighted multiple-instance learning

Nikolaos Pappas and Andrei Popescu-Belis. Explicit doc ument modeling through weighted multiple-instance learning. In JAIR, 2017

work page 2017

[15] [15]

Training region-based object detectors with online hard example mining

Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016

work page 2016

[16] [16]

Sirinukunwattana, S

K. Sirinukunwattana, S. E. A. Raza, Y. Tsang, D. R. J. Snea d, I. A. Cree, and N. M. Rajpoot. Locality sensitive deep learning for detecti on and classiﬁcation of nuclei in routine colon cancer histology images. In T-MI, 2016

work page 2016

[17] [17]

Y. Song, Q. Li, H. Huang, D. Feng, M. Chen, and W. Cai. Low di mensional representation of ﬁsher vectors for microscopy image class iﬁcation. In T-MI, 2017

work page 2017

[18] [18]

M. Sun, T. X. Han, , and A. Khodayari-Rostamabad. Multipl e instance learning convolutional neural networks for object recognition. In ICPR, 2016

work page 2016

[19] [19]

Deep Learning for Identifying Metastatic Breast Cancer

Dayong Wang, Aditya Khosla, Rishab Gargeya, Humayun Irs had, and Andrew H Beck. Deep learning for identifying metastatic breast canc er. arXiv preprint arXiv:1606.05718, 2016

work page internal anchor Pith review Pith/arXiv arXiv 2016

[20] [20]

Yan. Xu, J. Zhu, E. Chang, and Z. Tu. Multiple clustered in stance learning for histopathology cancer image classiﬁcation, segmentation and clustering. In CVPR, June 2012

work page 2012

[21] [21]

Xu, Tao Mo, Qiwei Feng, Peilin Zhong, Maode Lai, I Eri c, and Chao Chang

Yan. Xu, Tao Mo, Qiwei Feng, Peilin Zhong, Maode Lai, I Eri c, and Chao Chang. Deep learning of feature representation with multiple inst ance learning for medical image analysis. In ICASSP, 2014

work page 2014

[22] [22]

Deep adaptive fe ature embedding with local sample distributions for person re-identiﬁcati on

Lin Wu, Yang Wang, Junbin Gao, and Xue Li. Deep adaptive fe ature embedding with local sample distributions for person re-identiﬁcati on. Pattern Recognition, 73:275–288, 2018

work page 2018

[23] [23]

Deep attention- based spatially recursive networks for ﬁne-grained visual recognition

Lin Wu, Yang Wang, Xue Li, and Junbin Gao. Deep attention- based spatially recursive networks for ﬁne-grained visual recognition. IEEE Transactions on Cy- bernetics, 49(5):1791–1802, 2019

work page 2019

[24] [24]

Cycle-consistent deep g enerative hashing for cross-modal retrieval

Lin Wu, Yang Wang, and Ling Shao. Cycle-consistent deep g enerative hashing for cross-modal retrieval. IEEE Transactions on Image Processing , 28(4):1602–1612, 2019

work page 2019