Re-se-arch

Re-se-arch: arch, search and re-search, the three layers of research.

Intelligent systems work by putting parts together: objects into scenes, lines and planes into 3D structure, modules into networks, evidence into decisions. Whatever the system, three operations recur. It must compose parts into wholes, select among alternatives (including choosing none), and remember what it has learned while it keeps learning. We study these operations through grammars of parts, alternatives and compositions (the And-Or graphs we have developed since 2011) and through learning methods that infer such structure from data instead of fixing it by hand.

Each research area below is one chapter of that story, tagged with the operation it studies. Compose: how structure is built, in network architectures, in the geometry of scenes and in generated images. Select: how a model chooses, and how that choice can be steered or protected, from a classifier’s ranked predictions to the relations a graph network trusts. Remember: how a model keeps what it has learned while it adapts to new tasks. The question that ties them together, and drives our current work, is how to keep composition, selection and memory reliable as models scale and their tasks grow longer: what we call coherent computing.

Research areas

Compose Compositional and grammar-guided models: Can the structure of a model (which parts, which connections, which groupings) be learned from data instead of fixed by hand?
Compose Structured geometry: wireframes, planes and 3D: How can a network that predicts dense pixel maps reliably recover sparse geometric structure, such as line segments, wireframes, planes and 3D boxes, from images?
Compose Controllable generation: How can a generator be told where each object goes and what each one looks like, so a user can move, relabel or restyle one object while the rest of the image stays fixed?
Select Adversarial attack and robustness: How much control can a tiny, carefully chosen change to an image give an attacker over a classifier, down to dictating its whole ranked list of top predictions, and what do such attacks reveal about what the model has learned?
Select Graph neural networks and self-supervised learning: How can a graph neural network learn useful node representations when the edges cannot be fully trusted, because connected nodes are often unlike each other or the observed edges are noisy and incomplete?
Remember Continual and efficient adaptation: When a network must learn new tasks one after another, how should it decide which parts to reuse, which to adjust cheaply and which to add, without forgetting earlier tasks or wasting parameters?

Our research has been generously supported by ARO, NSF, IARPA, the Laboratory for Analytic Sciences (LAS), Sony and Salesforce (see Sponsors).

Compositional and grammar-guided models

Compose Can the structure of a model (which parts, which connections, which groupings) be learned from data instead of fixed by hand?

An And-Or Graph (AOG) stores a huge space of possible structures as one compact graph: an And-node composes parts, an Or-node chooses among alternative decompositions, and terminal nodes ground the parts in image data. Because sub-parts are shared, the best structure is found by exact dynamic programming rather than search; for example, a 10×12 grid with 2×2 parts encodes about 3.9 billion part configurations with only 5,209 And-nodes.

We first used AOGs to represent object parts, context and occlusion, then carried the same grammar into deep networks: over channel groups (AOGNets, where And means concatenate and Or means sum), over region-of-interest cells (iRCNN, so a detection comes with its parse tree), and, in reverse, as structure that emerges inside a generator under sparsity. A related line makes normalization and attention instance-specific: Attentive Normalization learns a mixture of affine transforms weighted per input, and Patch-to-Cluster attention (PaCa) lets image patches attend to a few learned clusters instead of to every other patch.

Overview slide of AOGNets. Left: an AOG building block with AND-nodes, OR-nodes and terminal nodes over the channels of an input feature map, and the node operations (splitting, concatenation, summation). Center: ImageNet top-1 accuracy against number of parameters for AOGNets, ResNets, ResNeXts, DenseNets and DPNs. Right: number of unique detectors per concept category from network dissection. Bottom left: activation heatmaps of ResNet and AOGNet models on example images. — AOGNets at a glance. The building block (left) composes channel groups with AND-nodes (concatenation) and OR-nodes (summation); the plot (center) shows ImageNet accuracy against model size, where AOGNets reach a given accuracy with fewer parameters. From our CVPR 2019 paper and its presentation slides.

How the idea developed

2016 · And-Or car model (IEEE TPAMI): an AOG that represents context and occlusion for car detection and viewpoint estimation.
2017 · AOGTracker (IEEE TPAMI): a full-structure AOG over part layouts, learned online while tracking, with a self-monitoring signal that triggers re-learning.
2019 · AOGNets (CVPR): the same grammar over channel groups; one generated building block contains ResNet, DenseNet, ResNeXt and DPN blocks as special cases.
2019 · iRCNN (ICCV): an AOG over the RoI grid, so each detection is explained by the parse tree that produced it.
2020 · Sparsified generator (CVPR): AND-OR structure is induced inside a generator by top-k sparsity rather than constructed.
2020 · Attentive Normalization (ECCV): replaces the single affine transform of a normalization layer with a per-input mixture, unifying normalization with channel attention.
2023 · PaCa-ViT (CVPR): patch-to-cluster attention in vision Transformers, with cost linear in the number of patches and cluster heatmaps as a by-product.

What it achieved

AOGNets: 22.28% ImageNet top-1 error with 11.9M parameters, against 23.0% for ResNet-152 with 60.2M. The saving is in parameters; at the largest model sizes the FLOPs are comparable to or above those of competing networks.
Attentive Normalization: reduces ResNet-50’s ImageNet top-1 error from 23.01% to 21.59% for 0.2M extra parameters and the same 4.09 GFLOPs, compared with 22.10% for Squeeze-and-Excitation.
AOGTracker: 60.10 one-pass success on the TB-100 benchmark against 46.17 for STRUCK, the best of the 29 trackers in that benchmark at the time.

More highlights

Overview slide of PaCa-ViT: patch-to-patch, patch-to-reduced-patch and patch-to-cluster attention compared, cluster heatmaps on example images, and three take-away points. — PaCa-ViT at a glance: image patches attend to a few learned clusters, whose heatmaps can be inspected directly. From our CVPR 2023 paper.

Structured geometry: wireframes, planes and 3D

Compose How can a network that predicts dense pixel maps reliably recover sparse geometric structure, such as line segments, wireframes, planes and 3D boxes, from images?

Segmentation networks are very good at predicting a value for every pixel, but line segments are sparse. The attraction field bridges the two: every pixel near a segment stores a small vector from which the segment can be recovered, so line detection becomes a dense regression problem. The first version (AFM) stored a 2D displacement, which pins down the line but not its endpoints; the holistic attraction field of HAWP stores four numbers that map each pixel into a segment-local frame, which recovers the segment exactly and makes its length irrelevant. That representation has stayed unchanged from 2020 to 2025 and was later lifted into 3D.

The same habit of finding a representation in which geometry becomes easy to predict runs through our work on planes and 3D objects: line segments become tokens that guide plane recovery, a single plane correspondence becomes a camera-pose hypothesis, and 3D detectors learn from auxiliary signals that are free at training time and discarded at inference. Much of this line is a long-running collaboration with Nan Xue (visiting PhD student in iVMCL, 2018–2020) and Gui-Song Xia’s group at Wuhan University.

Overview slide of HAWP. Top left: a line segment mapped into the local frame of a pixel by translation, rotation and scaling. Bottom left: an indoor image, its junction and line-segment proposals, and the parsed wireframe. Right: wireframes from AFM++, F-Clip, LETR and HAWP on building photos, and line detections from LSD, SOLD² and HAWPv3 on aerial images outside the training data. — HAWP at a glance. Each line segment is encoded by a closed-form 4D vector in the local frame of nearby pixels (top left), which turns wireframe parsing into an end-to-end trainable holistic attraction field (bottom left); the right side compares wireframes with other methods and shows HAWPv3 applied to images unlike its training data. From our HAWP papers (CVPR 2020, TPAMI 2023) and their presentation slides.

How the idea developed

2019 · AFM (CVPR): the 2D attraction field: line segment detection recast as dense region coloring.
2020 · HAWP (CVPR): the exact 4D holistic attraction field, with segments bound to junctions.
2021 · PlaneTR (ICCV): detected line segments become tokens that guide a Transformer recovering 3D planes.
2022 · MonoCon (AAAI): monocular 3D detection trained with projected 2D auxiliary contexts that are dropped at inference.
2022 · HoW-3D (3DV): a new task and dataset: recover the full 3D wireframe of an object, including its hidden part, from one image.
2023 · HAWPv2/v3 (IEEE TPAMI): a stronger detector and a self-supervised version of the attraction field.
2023 · NOPE-SAC (IEEE TPAMI): one plane correspondence per pose hypothesis, inside a RANSAC-style loop, for two-view planar reconstruction.
2023 · MonoXiver (ICCV): an oracle study shows that choosing among 3D proposals, not regressing them, is the bottleneck; a second stage re-selects.
2024 · NEAT (CVPR): the attraction field lifted into a 3D neural field, giving multi-view wireframe reconstruction without feature matching.
2024 · MvACon (CVPR): cluster-based context (from PaCa) added to 2D features before they are lifted to 3D in multi-view detectors.
2025 · ScaleLSD (CVPR): self-supervised line detection scaled to 10 million unlabeled images.

What it achieved

The representation is what matters: in HAWP, replacing the holistic attraction field with the earlier 2D field drops structural AP (sAP⁵) from 58.7 to 30.9 with everything else fixed. With it, HAWP reaches 65.7 msAP against 62.9 for a re-trained L-CNN, at 29.5 vs. 15.6 frames per second.
ScaleLSD: zero-shot structural repeatability (Rep-5) of 0.725 on YorkUrban and 0.666 on COCO, against 0.711 and 0.462 for HAWPv3 and 0.419 and 0.456 for the classical LSD detector.
MonoCon: 16.46 AP_3D on KITTI cars (moderate) at 25.8 ms per image, 1.44 points above GUPNet; the auxiliary contexts account for 8.49 points in ablation.
MvACon: on BEVFormer, local context changes NDS by −0.2 and global pooling by +0.3, while cluster context adds +1.1: the structure of the context, not context as such, is what helps.

More highlights

Overview slide of MvACon: detections on nuScenes by BEVFormer and by BEVFormer with MvACon, the multi-view 3D detection pipeline with MvACon between the 2D backbone and 2D-to-3D lifting, and three take-away points. — MvACon at a glance: cluster-based context added to 2D features before lifting improves query-based multi-view 3D detectors. From our CVPR 2024 paper.

Overview slide of NEAT: 3D wireframes reconstructed on DTU and BlendedMVS scenes, the NEAT field and junction distillation framework, and three take-away points. — NEAT at a glance: the attraction field lifted into a 3D neural field reconstructs 3D wireframes without cross-view matching. From our CVPR 2024 paper.

Overview slide of ScaleLSD: line segments detected by LSD, HAWPv3, DeepLSD and ScaleLSD on diverse images, the architecture and pseudo-label pipeline, and three take-away points. — ScaleLSD at a glance: self-supervised line detection trained on about 10 million unlabeled images. From our CVPR 2025 paper.

Controllable generation

Compose How can a generator be told where each object goes and what each one looks like, so a user can move, relabel or restyle one object while the rest of the image stays fixed?

A layout of labeled boxes is cheap to draw but too coarse to paint from directly. LostGAN factorizes generation as layout → per-object mask → image. Its instance-sensitive, layout-aware normalization (ISLA-Norm) gives every object its own normalization scale and shift, computed from the object’s class and its own style code, and places them in the image through that object’s predicted soft mask; the masks are learned without any mask labels. Because each object has its own style code and its own footprint, changing one object’s code or box changes only that object. Our later work applies the same principle inside excitation layers: a channel-only gate cannot say where, so spatially-adaptive modules output a full spatial modulation from separate “what” and “where” branches.

The results below are from the GAN era; diffusion models have since raised the bar for image quality, and the ideas of per-object normalization and learned layout masks carry over to them.

Overview slide of LostGAN. Top: a layout of labeled boxes (sky, snow, mountain, tree, person, skis) plus a style code is synthesized into an image. Bottom left: the layout-to-mask-to-image generator with ISLA-Norm. Right: snow scenes generated as object boxes are added one by one and as the person's box is moved, with three style codes per layout. — LostGAN at a glance. A layout says where to place what, and per-object style codes say how each object looks; the layout-to-mask-to-image generator with ISLA-Norm (bottom left) turns them into an image. On the right, adding boxes or moving the person’s box changes only those objects. From our LostGAN papers (ICCV 2019, TPAMI 2022) and their presentation slides.

How the idea developed

2019 · LostGAN (ICCV): layout → mask → image with ISLA-Norm and per-object style codes; reconfigurable layout and style.
2021 · Deep Consensus Learning (arXiv:2103.08475): adds a segmentation network as a third player trained on the generator’s own masks, closing the loop between synthesis and recognition (preprint).
2022 · LostGAN-V2 (IEEE TPAMI): multi-stage mask refinement, 256×256 and 512×512 images, and training with box labels on only half of the images.
2023 · SASM (ICIP): separate appearance and layout branches for generation from a single image.
2023 · SASE (ICIP): spatially-adaptive squeeze-excitation for generators trained on 100 to 1,000 images.

What it achieved

LostGAN (2019, COCO-Stuff 64×64): diversity score 0.35 against 0.15 for Layout2Im, with better image quality (FID 34.31 vs. 38.14).
LostGAN-V2 (COCO-Stuff 256×256): FID 42.55 against 65.95 for Grid2Im, which needs ground-truth masks. A detector finds objects in our generated images with 34.9 AP, against 32.3 for Grid2Im and 44.7 on real images.
SASE: as a drop-in module in FastGAN it improves FID on all 12 low-shot datasets tested (FFHQ: 44.3 → 39.59) with fewer parameters (28.5M vs. 29.1M), and the gain holds under a different augmentation scheme.

Adversarial attack and robustness

Select How much control can a tiny, carefully chosen change to an image give an attacker over a classifier, down to dictating its whole ranked list of top predictions, and what do such attacks reveal about what the model has learned?

Most attacks only flip the top-1 label. We ask for more: make the model rank chosen classes t₁ > t₂ > … > t_K above all others. That requirement is a set of linear inequalities on the model’s outputs, so the smallest perturbation that satisfies it can be found by constrained optimization rather than a hand-tuned loss. QuadAttacK solves it as a quadratic program on the network’s features; RisingAttacK solves it in image space by sequential quadratic programming, where each step combines the right singular vectors of the Jacobian from image to logits. When gradients are not available, our black-box attacks follow the geometry of the decision boundary instead: CGBA searches along a semicircle rather than a line, which still works when the boundary is curved. Attacks are also a diagnostic: an attack that preserves the matching cue still breaks stereo networks, which shows they were not really matching.

Overview slide of RisingAttacK. Left top: workflow of ordered top-K attacks, one constrained objective solved three ways (RisingAttacK in image space, QuadAttacK in feature space, AdvDistill with a surrogate loss). Left bottom: a redshank image attacked on ResNet-50 for top-K from 1 to 30, with the perturbation for each K. Right: three take-away points. — RisingAttacK at a glance. Every ordered top-K attack minimizes the perturbation subject to the same linear constraints on the logits; RisingAttacK solves them directly in image space. Below, the attacked images look unchanged while the required perturbation grows with K. From our ICML 2025 paper.

How the idea developed

2020 · Adversarial distillation (CVPR Workshops): defines ordered top-K targeted attacks and a first method for them.
2022 · Robust stereo (NeurIPS): uses a cue-preserving attack as a probe, then builds stereo matching on a non-learned matching cost that resists it.
2023 · CGBA (ICCV): a label-only black-box attack that searches the boundary along a semicircle.
2023 · QuadAttacK (NeurIPS): ordered top-K attacks as a quadratic program in feature space; K up to 20.
2025 · GSBA^K (ICLR): the first score-based black-box attack on top-K predictions (unordered sets).
2025 · RisingAttacK (ICML): ordered top-K attacks solved directly in image space; K up to 30 on ResNet, DenseNet and vision Transformers.

What it achieved

RisingAttacK: across K ∈ {1, 5, …, 30} on four ImageNet models, its figure of merit is more than twice QuadAttacK’s in most cases, and it succeeds at K = 30 where QuadAttacK fails on DenseNet-121, ViT-B and DeiT-B.
CGBA: median ℓ₂ perturbation of 0.86 on ResNet-50/ImageNet after 10,000 queries (untargeted), against 1.58 for SurFree and 2.49 for HSJA.
GSBA^K: 78.8% success for targeted top-3 attacks on ResNet-50 with 40,000 queries, against 31.8% for the previous score-based top-K attack.
Robust stereo: trained on synthetic data only, its error on KITTI 2015 (Bad-1.0) is 25.95 against 42.12 for GANet, and under attack its end-point error moves from 0.36 to 0.88 while PSMNet’s moves from 0.28 to 29.05.

Graph neural networks and self-supervised learning

Select How can a graph neural network learn useful node representations when the edges cannot be fully trusted, because connected nodes are often unlike each other or the observed edges are noisy and incomplete?

Most graph neural networks assume homophily: connected nodes are similar, so averaging over neighbors helps. Many real graphs are heterophilic, and most mix both. HarmonyGNNs lets each node decide how much to trust its neighborhood: it encodes every node as a few tokens (its own features, and features aggregated over one and over several hops) and a small Transformer weighs them per node. It learns without labels, from a student network that sees a masked graph and predicts the embeddings of a slowly updated teacher. ReCoG goes one step further and learns the graph itself: in multimodal graphs, where nodes carry text and images, it re-scores and adds edges from cross-modal evidence, and refines the graph and the node representations jointly, so a better graph yields better embeddings and vice versa.

Overview slide of ReCoG. Left top: the framework, a multimodal graph refined from text and image evidence while cross-modal message passing updates text and vision features, trained end to end. Left bottom: similarity distributions on RedditM showing that edges added by the refiner are far more similar than disconnected node pairs. Right: three take-away points. — ReCoG at a glance. The graph and the node representations are refined together, and the edges the refiner adds connect nodes whose text and images agree. From our ReCoG preprint (arXiv:2608.22786).

How the idea developed

2026 · HarmonyGNNs (ICLR): self-supervised node representations that work across homophilic and heterophilic graphs without being told which kind of graph it is.
2026 · ReCoG (arXiv:2608.22786): graph structure and multimodal node representations learned jointly, end to end (preprint).

What it achieved

HarmonyGNNs (linear probing): 75.86% on Roman-Empire against 66.26% for MUSE, and 93.24% on Texas against 86.12% for S3GCL, while matching the best self-supervised methods on homophilic graphs such as Cora (84.82% vs. 84.20%).
ReCoG: 90.98% node-classification accuracy on Goodreads against 84.16% for PLANET. On Ele-Fashion, removing the graph refiner drops accuracy from 88.56% to 87.09%, and removing cross-modal message passing drops it to 86.83%: the two parts are complementary.

More highlights

Overview slide of HarmonyGNNs: the teacher-student framework with joint structural node encoding, a bar chart of GCN, GAT and MLP accuracy on heterophilic and homophilic synthetic graphs, and three take-away points. — HarmonyGNNs at a glance: each node decides how much to trust its neighborhood, learned without labels. From our ICLR 2026 paper.

Continual and efficient adaptation

Remember When a network must learn new tasks one after another, how should it decide which parts to reuse, which to adjust cheaply and which to add, without forgetting earlier tasks or wasting parameters?

We treat a network’s structure and its parameters as separate decisions. For each layer and each new task, a search chooses whether to reuse the layer as is, adapt it with a small add-on, add a new layer, or (in CHEEM) skip the block, with a penalty on added parameters or compute. Reused layers stay frozen, so earlier tasks are protected by construction rather than by a penalty term. For adapting large pretrained models cheaply, WeGeFT generates the weight update from the pretrained weights themselves through two small matrices shared across layers, so adaptation needs very few new parameters while keeping the compute and memory cost of LoRA.

Overview slide of CHEEM. Left: the method, a Vision Transformer whose blocks are reused, adapted, newly added or skipped for each task, with an external memory of task centroids that infers the task at test time. Right top: the models learned for four tasks on ViT-Base. Right bottom: three take-away points. — CHEEM at a glance. For each new task, CHEEM searches over four operations per block, so tasks close to ImageNet reuse most of the network and distant tasks adapt and add more. From our CVPR 2026 paper.

How the idea developed

2019 · Learn to Grow (ICML): per-layer reuse / adapt / new, chosen by differentiable architecture search, for convolutional networks.
2025 · WeGeFT (ICML): fine-tuning updates generated from the pretrained weights, for large language models and vision Transformers.
2026 · CHEEM (CVPR): adds Skip and a task-similarity-guided search, for Vision Transformers in exemplar-free class-incremental learning.

What it achieved

Learn to Grow: 77.68% average accuracy over the 10 tasks of the Visual Domain Decathlon with 14.46M parameters, against 76.02% for residual adapters (12.50M) and 71.70% for ten separate models (58.96M).
CHEEM: 85.9% average accuracy over 12 tasks (MTIL, ImageNet-pretrained ViT-Base) with 1.7% forgetting, against 81.6% for S-Prompts, the strongest prompting baseline, and 88.1% for fine-tuning a separate model per task.
WeGeFT: fine-tuning Llama 2 (7B) for math reasoning, 54.89% on GSM8K while training 0.068% of the parameters, on par with full fine-tuning (54.20%) and above LoRA-GA (53.60%, 0.297%) and DoRA (53.07%, 0.317%).

More highlights

Overview slide of WeGeFT: LoRA and WeGeFT compared as diagrams, accuracy against trainable parameters on commonsense, code and math tasks, and three take-away points. — WeGeFT at a glance: fine-tuning updates generated from the pretrained weights, with very few trainable parameters. From our ICML 2025 paper.

Publications

2026

Conceal, Reconstruct, Jailbreak: Exploiting the Reconstruction–Concealment Tradeoff in MLLMs. Md Farhamdur Reza, Richeng Jin, Tianfu Wu and Huaiyu Dai. NeurIPS, 2026.
AI-Based Reverberation Detection to Enhance HIFU Delivery Safety Based on B-Mode Ultrasound Images with Overlays. Ralf Seip, Tianfu Wu, Xiangming Xue, Xiaoning Jiang, Patrick Fitzgerald, Darrel May, Narendra Sanghvi and Richard Yang. IUS, 2026.
CHEEM: Continual Learning by Reuse, New, Adapt and Skip– A Hierarchical Exploration-Exploitation Approach. Chinmay Savadikar, Michelle Dai and Tianfu Wu. CVPR, 2026.
HarmonyGNNs: Harmonizing Heterophily and Homophily in GNNs via Joint Structural Node Encoding and Self-Supervised Learning. Rui Xue and Tianfu Wu. ICLR, 2026.
ReCoG: Reciprocal Co-Evolution for Multimodal Graph Learning. Rui Xue and Tianfu Wu. arXiv:2608.22786, 2026.
ROPE-MLA: Row-Access Optimized Processing Element Machine Learning Accelerator. Prasanth Prabu Ravichandiran, Paul Franzon, William Davis, Franc Brglez and Tianfu Wu. 3DIC, 2026.
ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents. Yuanzheng Zhu, Mingyu Zhao, Chinmay Savadikar, Han Li, Shuang Xie, Alberto Castelo, Tianfu Wu and Lingyun Wang. NeurIPS Evaluations and Datasets, 2026.
ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents. Chinmay Savadikar, Mingyu Zhao, Yuanzheng Zhu, Han Li, Shuang Xie, Alberto Castelo, Tianfu Wu and Lingyun Wang. COLM Learning from Situated and Embodied Interaction Workshop, 2026.

2025

Adversarial Perturbations Are Formed by Iteratively Learning Linear Combinations of the Right Singular Vectors of the Adversarial Jacobian. Thomas Paniagua, Chinmay Savadikar and Tianfu Wu. ICML, 2025.
DiffMesh: A Motion-Aware Diffusion Framework for Human Mesh Recovery from Videos. Ce Zheng, Xianpeng Liu, Qucheng Peng, Tianfu Wu, Pu Wang and Chen Chen. WACV, 2025.
Sign-Based Gradient Descent With Heterogeneous Data: Convergence and Byzantine Resilience. Richeng Jin, Yuding Liu, Yufan Huang, Xiaofan He, Tianfu Wu and Huaiyu Dai. IEEE TNNLS, 2025.
WeGeFT: Weight-Generative Fine-Tuning for Multi-Faceted Efficient Adaptation of Large Models. Chinmay Savadikar, Xi Song and Tianfu Wu. ICML, 2025.
GSBA^K: top-K Geometric Score-based Black-box Attack. Md Farhamdur Reza, Richeng Jin, Tianfu Wu and Huaiyu Dai. ICLR, 2025.
ScaleLSD: Scalable Deep Line Segment Detection Streamlined. Zeran Ke, Bin Tan, Xianwei Zheng, Yujun Shen, Tianfu Wu and Nan Xue. CVPR, 2025.

2024

A Pixel Is Worth More Than One 3D Gaussians in Single-View 3D Reconstruction. Jianghao Shen, Nan Xue and Tianfu Wu. arXiv:2405.20310, 2024.
Enhancing Approximate Message Passing via Diffusion Models Towards On-Device Intelligence. Philip Dai, Kai Yue, Richeng Jin, Matt Tianfu Wu and Kaiqi Xiong. ICC Workshops, 2024.
High-Speed Receiver Transient Modeling with Generative Adversarial Networks. Priyank Kashyap, Andries Deroo, Dror Baron, Chau-Wai Wong, Tianfu Wu and Paul D. Franzon. MDTS, 2024.
Multi-View Attentive Contextualization for Multi-View 3D Object Detection. Xianpeng Liu, Ce Zheng, Ming Qian, Nan Xue, Chen Chen, Zhebin Zhang, Chen Li and Tianfu Wu. CVPR, 2024.
NEAT: Distilling 3D Wireframes from Neural Attraction Fields. Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu and Yujun Shen. CVPR, 2024.

2023

Holistically-Attracted Wireframe Parsing: From Supervised to Self-Supervised Learning. Nan Xue, Tianfu Wu, Song Bai, Fu-Dong Wang, Gui-Song Xia, Liangpei Zhang and Philip H. S. Torr. IEEE TPAMI, 2023.
Implicit Bayes Adaptation: A Collaborative Transport Approach. Bo Jiang, Hamid Krim, Tianfu Wu and Derya Cansever. Signal Processing, 2023.
Learning Spatially-Adaptive Squeeze-Excitation Networks for Few Shot Image Synthesis. Jianghao Shen and Tianfu Wu. ICIP, 2023.
Learning Spatially-Adaptive Style-Modulation Networks for Single Image Synthesis. Jianghao Shen and Tianfu Wu. ICIP, 2023.
Level-S²fM: Structure from Motion on Neural Level Set of Implicit Surfaces. Yuxi Xiao, Nan Xue, Tianfu Wu and Gui-Song Xia. CVPR, 2023.
Monocular 3D Object Detection with Bounding Box Denoising in 3D by Perceiver. Xianpeng Liu, Ce Zheng, Kelvin Cheng, Nan Xue, Guo-Jun Qi and Tianfu Wu. ICCV, 2023.
PaCa-ViT: Learning Patch-to-Cluster Attention in Vision Transformers. Ryan Grainger, Thomas Paniagua, Xi Song, Naresh Cuntoor, Mun Wai Lee and Tianfu Wu. CVPR, 2023.
QuadAttacK: A Quadratic Programming Approach to Learning Ordered Top-K Adversarial Attacks. Thomas Paniagua, Ryan Grainger and Tianfu Wu. NeurIPS, 2023.
Thermal Estimation for 3D-ICs Through Generative Networks. Priyank Kashyap, Prasanth Prabu Ravichandiran, Lee Wang, Dror Baron, Chau-Wai Wong, Tianfu Wu and Paul D. Franzon. 3DIC, 2023.
CGBA: Curvature-aware Geometric Black-box Attack. Md Farhamdur Reza, Ali Rahmati, Tianfu Wu and Huaiyu Dai. ICCV, 2023.
NOPE-SAC: Neural One-Plane RANSAC for Sparse-View Planar 3D Reconstruction. Bin Tan, Nan Xue, Tianfu Wu and Gui-Song Xia. IEEE TPAMI, 2023.

2022

HoW-3D: Holistic 3D Wireframe Perception from a Single Image. Wenchao Ma, Bin Tan, Nan Xue, Tianfu Wu, Xianwei Zheng and Gui-Song Xia. 3DV, 2022.
Learning Auxiliary Monocular Contexts Helps Monocular 3D Object Detection. Xianpeng Liu, Nan Xue and Tianfu Wu. AAAI, 2022.
Learning Layout and Style Reconfigurable GANs for Controllable Image Synthesis. Wei Sun and Tianfu Wu. IEEE TPAMI, 2022.
Learning Local-Global Contextual Adaptation for Multi-Person Pose Estimation. Nan Xue, Tianfu Wu, Gui-Song Xia and Liangpei Zhang. CVPR, 2022.
Preliminary Evaluation of a System with On-Body and Aerial Sensors for Monitoring Working Dogs. Marc Foster, Tianfu Wu, David L. Roberts and Alper Bozkurt. Sensors, 2022.
Refining Self-Supervised Learning in Imaging: Beyond Linear Metric. Bo Jiang, Hamid Krim, Tianfu Wu and Derya Cansever. ICIP, 2022.
Revisiting Non-Parametric Matching Cost Volumes for Robust and Generalizable Stereo Matching. Kelvin Cheng, Tianfu Wu and Christopher G. Healey. NeurIPS, 2022.
RxGAN: Modeling High-Speed Receiver through Generative Adversarial Networks. Priyank Kashyap, Archit Gajjar, Yongjin Choi, Chau-Wai Wong, Dror Baron, Tianfu Wu, Chris Cheng and Paul D. Franzon. MLCAD, 2022.

2021

A Bottom-Up and Top-Down Integration Framework for Online Object Tracking. Meihui Li, Lingbing Peng, Tianfu Wu and Zhenming Peng. IEEE TMM, 2021.
Deep Consensus Learning. Wei Sun and Tianfu Wu. arXiv:2103.08475, 2021.
Event driven sensor fusion. Siddharth Roheda, Hamid Krim, Zhi-Quan Luo and Tianfu Wu. Signal Processing, 2021.
Learning Regional Attraction for Line Segment Detection. Nan Xue, Song Bai, Fu-Dong Wang, Gui-Song Xia, Tianfu Wu, Liangpei Zhang and Philip H. S. Torr. IEEE TPAMI, 2021.
PlaneTR: Structure-Guided Transformers for 3D Plane Recovery. Bin Tan, Nan Xue, Song Bai, Tianfu Wu and Gui-Song Xia. ICCV, 2021.

2020

Attentive Normalization. Xilai Li, Wei Sun and Tianfu Wu. ECCV, 2020.
Holistically-Attracted Wireframe Parsing. Nan Xue, Tianfu Wu, Song Bai, Fudong Wang, Gui-Song Xia, Liangpei Zhang and Philip H. S. Torr. CVPR, 2020.
Inducing Hierarchical Compositional Model by Sparsifying Generator Network. Xianglei Xing, Tianfu Wu, Song-Chun Zhu and Ying Nian Wu. CVPR, 2020.
Learning Ordered Top-k Adversarial Attacks via Adversarial Distillation. Zekun Zhang and Tianfu Wu. CVPR Workshops, 2020.
Local Clustering with Mean Teacher for Semi-supervised learning. Zexi Chen, Benjamin Dutton, Bharathkumar Ramachandra, Tianfu Wu and Ranga Raju Vatsavai. ICPR, 2020.

2019

AOGNets: Compositional Grammatical Architectures for Deep Learning. Xilai Li, Xi Song and Tianfu Wu. CVPR, 2019.
Image Synthesis From Reconfigurable Layout and Style. Wei Sun and Tianfu Wu. ICCV, 2019.
Jointly social grouping and identification in visual dynamics with causality-induced hierarchical Bayesian model. Zhao Xie, Tianfu Wu, Xingming Yang, Luming Zhang and Kewei Wu. JVCIR, 2019.
Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting. Xilai Li, Yingbo Zhou, Tianfu Wu, Richard Socher and Caiming Xiong. ICML, 2019.
Learning Attraction Field Representation for Robust Line Segment Detection. Nan Xue, Song Bai, Fudong Wang, Gui-Song Xia, Tianfu Wu and Liangpei Zhang. CVPR, 2019.
Learning Spatial Pyramid Attentive Pooling in Image Synthesis and Image-to-Image Translation. Wei Sun and Tianfu Wu. arXiv:1901.06322, 2019.
Pose Guided Fashion Image Synthesis Using Deep Generative Model. Wei Sun, Jawadul H. Bappy, Shanglin Yang, Yi Xu, Tianfu Wu and Hui Zhou. arXiv:1906.07251, 2019.
Towards Interpretable Object Detection by Unfolding Latent Structures. Tianfu Wu and Xi Song. ICCV, 2019.

2018

Auto-Context R-CNN. Bo Li, Tianfu Wu, Lun Zhang and Rufeng Chu. arXiv:1807.02842, 2018.
Decision Level Fusion: An Event Driven Approach. Siddharth Roheda, Hamid Krim, Zhi-Quan Luo and Tianfu Wu. Signal Processing, 2018.
High Resolution Face Completion with Multiple Controllable Attributes via Fully End-to-End Progressive Generative Adversarial Networks. Zeyuan Chen, Shaoliang Nie, Tianfu Wu and Christopher G. Healey. arXiv:1801.07632, 2018.
Neural Abstract Style Transfer for Chinese Traditional Painting. Bo Li, Caiming Xiong, Tianfu Wu, Yu Zhou, Lun Zhang and Rufeng Chu. ACCV, 2018.
Scene-Centric Joint Parsing of Cross-View Videos. Hang Qi, Yuanlu Xu, Tao Yuan, Tianfu Wu and Song-Chun Zhu. AAAI, 2018.
ARCHER: Aggressive Rewards to Counter bias in Hindsight Experience Replay. Sameera Lanka and Tianfu Wu. arXiv:1809.02070, 2018.

2017

Joint Parsing of Cross-view Scenes with Spatio-temporal Semantic Parse Graphs. Hang Qi, Yuanlu Xu, Tao Yuan, Tianfu Wu and Song-Chun Zhu. arXiv:1709.05436, 2017.
Online Object Tracking, Learning and Parsing with And-Or Graphs. Tianfu Wu, Yang Lu and Song-Chun Zhu. IEEE TPAMI, 2017.
Zero-Shot Learning Posed as a Missing Data Problem. Bo Zhao, Botong Wu, Tianfu Wu and Yizhou Wang. ICCV Workshops, 2017.

2016

Deep Learning for Remote Sensing Image Understanding. Liangpei Zhang, Gui-Song Xia, Tianfu Wu, Liang Lin and Xue-Cheng Tai. Journal of Sensors, 2016.
Face Detection with End-to-End Integration of a ConvNet and a 3D Model. Yunzhu Li, Benyuan Sun, Tianfu Wu and Yizhou Wang. ECCV, 2016.
Learning And-Or Model to Represent Context and Occlusion for Car Detection and Viewpoint Estimation. Tianfu Wu, Bo Li and Song-Chun Zhu. IEEE TPAMI, 2016.
Monocular Depth-Ordering Reasoning with Occlusion Edge Detection and Couple Layers Inference. Anlong Ming, Tianfu Wu, Jianxiang Ma, Fang Sun and Yu Zhou. IEEE Intelligent Systems, 2016.