本文首發(fā)極市平臺公眾號，轉(zhuǎn)載請獲得授權(quán)并標(biāo)明出處。

ECCV 2022 論文分方向整理目前在極市社區(qū)持續(xù)更新中，已累計更新了54篇，項目地址：https://github.com/extreme-assistant/ECCV2022-Paper-Code-Interpretation

以下是本周更新的 ECCV 2022 論文，包含檢測，分割，圖像處理，視頻理解，神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)設(shè)計，無監(jiān)督學(xué)習(xí)，遷移學(xué)習(xí)等方向。

論文合集打包下載地址：https://www.cvmart.net/community/detail/6592

– 檢測 – 分割 – 圖像處理 – 視頻處理 – 圖像、視頻檢索與理解 – 估計- 目標(biāo)跟蹤 – 文本檢測與識別 – GAN/生成式/對抗式 – 神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)設(shè)計 – 數(shù)據(jù)處理 – 模型訓(xùn)練/泛化 – 模型壓縮 – 模型評估 – 半監(jiān)督學(xué)習(xí)/自監(jiān)督學(xué)習(xí) – 多模態(tài)/跨模態(tài)學(xué)習(xí) – 小樣本學(xué)習(xí) – 強化學(xué)習(xí)

檢測

2D目標(biāo)檢測

[1] Point-to-Box Network for Accurate Object Detection via Single Point Supervision (通過單點監(jiān)督實現(xiàn)精確目標(biāo)檢測的點對盒網(wǎng)絡(luò))paper：https://arxiv.org/abs/2207.06827code：https://github.com/ucas-vg/p2bnet

[2] You Should Look at All Objects (您應(yīng)該查看所有物體)paper：https://arxiv.org/abs/2207.07889code：https://github.com/charlespikachu/yslao

[3] Adversarially-Aware Robust Object Detector (對抗性感知魯棒目標(biāo)檢測器)paper：https://arxiv.org/abs/2207.06202code：https://github.com/7eu7d7/robustdet

3D目標(biāo)檢測

[1] Rethinking IoU-based Optimization for Single-stage 3D Object Detection (重新思考基于 IoU 的單階段 3D 對象檢測優(yōu)化)paper：https://arxiv.org/abs/2207.09332

人物交互檢測

[1] Towards Hard-Positive Query Mining for DETR-based Human-Object Interaction Detection (面向基于 DETR 的人機交互檢測的硬性查詢挖掘)paper：https://arxiv.org/abs/2207.05293 code：https://github.com/muchhair/hqm

圖像異常檢測

[1] DICE: Leveraging Sparsification for Out-of-Distribution Detection (DICE：利用稀疏化進(jìn)行分布外檢測)paper：https://arxiv.org/abs/2111.09805code：https://github.com/deeplearning-wisc/dice

分割

實例分割

[1] Box-supervised Instance Segmentation with Level Set Evolution (具有水平集進(jìn)化的框監(jiān)督實例分割)paper：https://arxiv.org/abs/2207.09055

[2] OSFormer: One-Stage Camouflaged Instance Segmentation with Transformers (OSFormer：使用 Transformers 進(jìn)行單階段偽裝實例分割)paper：https://arxiv.org/abs/2207.02255 code：https://github.com/pjlallen/osformer

語義分割

[1] 2DPASS: 2D Priors Assisted Semantic Segmentation on LiDAR Point Clouds (2DPASS：激光雷達(dá)點云上的二維先驗輔助語義分割)paper：https://arxiv.org/abs/2207.04397 code：https://github.com/yanx27/2dpass

視頻目標(biāo)分割

[1] Learning Quality-aware Dynamic Memory for Video Object Segmentation (視頻對象分割的學(xué)習(xí)質(zhì)量感知動態(tài)內(nèi)存)paper：https://arxiv.org/abs/2207.07922code：https://github.com/workforai/qdmn

圖像處理

超分辨率

[1] Dynamic Dual Trainable Bounds for Ultra-low Precision Super-Resolution Networks (超低精度超分辨率網(wǎng)絡(luò)的動態(tài)雙可訓(xùn)練邊界)paper：https://arxiv.org/abs/2203.03844 code：https://github.com/zysxmu/ddtb

圖像去噪

[1] Deep Semantic Statistics Matching (D2SM) Denoising Network (深度語義統(tǒng)計匹配（D2SM）去噪網(wǎng)絡(luò))paper：https://arxiv.org/abs/2207.09302

圖像復(fù)原/圖像增強/圖像重建

[1] Semantic-Sparse Colorization Network for Deep Exemplar-based Colorization (用于基于深度示例的著色的語義稀疏著色網(wǎng)絡(luò))paper：https://arxiv.org/abs/2112.01335

[2] Geometry-aware Single-image Full-body Human Relighting (幾何感知單圖像全身人體重新照明)paper：https://arxiv.org/abs/2207.04750

[3] Multi-Modal Masked Pre-Training for Monocular Panoramic Depth Completion (單目全景深度補全的多模態(tài)蒙面預(yù)訓(xùn)練)paper：https://arxiv.org/abs/2203.09855

[4] PanoFormer: Panorama Transformer for Indoor 360 Depth Estimation (PanoFormer：用于室內(nèi) 360 深度估計的全景變壓器)paper：https://arxiv.org/abs/2203.09283

[5] SESS: Saliency Enhancing with Scaling and Sliding (SESS：通過縮放和滑動增強顯著性)paper：https://arxiv.org/abs/2207.01769

[6] RigNet: Repetitive Image Guided Network for Depth Completion (RigNet：用于深度補全的重復(fù)圖像引導(dǎo)網(wǎng)絡(luò))paper：https://arxiv.org/abs/2107.13802

圖像外推(Image Outpainting)

[1] Outpainting by Queries (通過查詢進(jìn)行外包)paper：https://arxiv.org/abs/2207.05312 code：https://github.com/kaiseem/queryotr

風(fēng)格遷移(Style Transfer)

[1] CCPL: Contrastive Coherence Preserving Loss for Versatile Style Transfer (CCPL：通用風(fēng)格遷移的對比相干性保留損失)paper：https://arxiv.org/abs/2207.04808 code：https://github.com/JarrentWu1031/CCPL

視頻處理(Video Processing)

[1] Improving the Perceptual Quality of 2D Animation Interpolation (提高二維動畫插值的感知質(zhì)量)paper：https://arxiv.org/abs/2111.12792code：https://github.com/shuhongchen/eisai-anime-interpolator

[2] Real-Time Intermediate Flow Estimation for Video Frame Interpolation (視頻幀插值的實時中間流估計)paper：https://arxiv.org/abs/2011.06294 code：https://github.com/MegEngine/arXiv2020-RIFE

圖像、視頻檢索與理解

動作識別

[1] ReAct: Temporal Action Detection with Relational Queries (ReAct：使用關(guān)系查詢的時間動作檢測)paper：https://arxiv.org/abs/2207.07097code：https://github.com/sssste/react

[2] Hunting Group Clues with Transformers for Social Group Activity Recognition (用Transformers尋找群體線索用于社會群體活動識別)paper：https://arxiv.org/abs/2207.05254

視頻理解

[1] GraphVid: It Only Takes a Few Nodes to Understand a Video (GraphVid：只需幾個節(jié)點即可理解視頻)paper：https://arxiv.org/abs/2207.01375

[2] Deep Hash Distillation for Image Retrieval (用于圖像檢索的深度哈希蒸餾)paper：https://arxiv.org/abs/2112.08816code：https://github.com/youngkyunjang/deep-hash-distillation

視頻檢索(Video Retrieval)

[1] TS2-Net: Token Shift and Selection Transformer for Text-Video Retrieval (TS2-Net：用于文本視頻檢索的令牌移位和選擇轉(zhuǎn)換器)paper：https://arxiv.org/abs/2207.07852code：https://github.com/yuqi657/ts2_net

[2] Lightweight Attentional Feature Fusion: A New Baseline for Text-to-Video Retrieval (輕量級注意力特征融合：文本到視頻檢索的新基線)paper：https://arxiv.org/abs/2112.01832

估計

位姿估計

[1] Category-Level 6D Object Pose and Size Estimation using Self-Supervised Deep Prior Deformation Networks (使用自監(jiān)督深度先驗變形網(wǎng)絡(luò)的類別級 6D 對象姿勢和大小估計)paper：https://arxiv.org/abs/2207.05444 code：https://github.com/jiehonglin/self-dpdn

深度估計

[1] Physical Attack on Monocular Depth Estimation with Optimal Adversarial Patches (使用最優(yōu)對抗補丁對單目深度估計進(jìn)行物理攻擊)paper：https://arxiv.org/abs/2207.04718

目標(biāo)跟蹤

[1] Towards Grand Unification of Object Tracking (邁向目標(biāo)跟蹤的大統(tǒng)一)paper：https://arxiv.org/abs/2207.07078code：https://github.com/masterbin-iiau/unicorn

文本檢測與識別

[1] Dynamic Low-Resolution Distillation for Cost-Efficient End-to-End Text Spotting (用于經(jīng)濟(jì)高效的端到端文本識別的動態(tài)低分辨率蒸餾)paper：https://arxiv.org/abs/2207.06694code：https://github.com/hikopensource/davar-lab-ocr

GAN/生成式/對抗式

[1] Eliminating Gradient Conflict in Reference-based Line-Art Colorization (消除基于參考的藝術(shù)線條著色中的梯度沖突)paper：https://arxiv.org/abs/2207.06095code：https://github.com/kunkun0w0/sga

[2] WaveGAN: Frequency-aware GAN for High-Fidelity Few-shot Image Generation (WaveGAN：用于高保真少鏡頭圖像生成的頻率感知 GAN)paper：https://arxiv.org/abs/2207.07288code：https://github.com/kobeshegu/eccv2022_wavegan

[3] FakeCLR: Exploring Contrastive Learning for Solving Latent Discontinuity in Data-Efficient GANs (FakeCLR：探索對比學(xué)習(xí)以解決數(shù)據(jù)高效 GAN 中的潛在不連續(xù)性)paper：https://arxiv.org/abs/2207.08630code：https://github.com/iceli1007/fakeclr

[4] UniCR: Universally Approximated Certified Robustness via Randomized Smoothing (UniCR：通過隨機平滑獲得普遍近似的認(rèn)證魯棒性)paper：https://arxiv.org/abs/2207.02152

神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)設(shè)計

神經(jīng)網(wǎng)絡(luò)架構(gòu)搜索(NAS)

[1] ScaleNet: Searching for the Model to Scale (ScaleNet：搜索要擴展的模型)paper：https://arxiv.org/abs/2207.07267code：https://github.com/luminolx/scalenet

[2] Ensemble Knowledge Guided Sub-network Search and Fine-tuning for Filter Pruning (集成知識引導(dǎo)的子網(wǎng)絡(luò)搜索和過濾器修剪微調(diào))paper：https://arxiv.org/abs/2203.02651 code：https://github.com/sseung0703/ekg

[3] EAGAN: Efficient Two-stage Evolutionary Architecture Search for GANs (EAGAN：GAN 的高效兩階段進(jìn)化架構(gòu)搜索)paper：https://arxiv.org/abs/2111.15097 code：https://github.com/marsggbo/EAGAN

數(shù)據(jù)處理

歸一化

[1] Fine-grained Data Distribution Alignment for Post-Training Quantization (訓(xùn)練后量化的細(xì)粒度數(shù)據(jù)分布對齊)paper：https://arxiv.org/abs/2109.04186 code：https://github.com/zysxmu/fdda

模型訓(xùn)練/泛化

噪聲標(biāo)簽

[1] Learning with Noisy Labels by Efficient Transition Matrix Estimation to Combat Label Miscorrection (通過有效的轉(zhuǎn)移矩陣估計學(xué)習(xí)噪聲標(biāo)簽以對抗標(biāo)簽錯誤校正)paper：https://arxiv.org/abs/2111.14932

模型壓縮

知識蒸餾

[1] Knowledge Condensation Distillation (知識濃縮蒸餾)paper：https://arxiv.org/abs/2207.05409 code：https://github.com/dzy3/kcd)

模型評估

[1] Hierarchical Latent Structure for Multi-Modal Vehicle Trajectory Forecasting (多模式車輛軌跡預(yù)測的分層潛在結(jié)構(gòu))paper：https://arxiv.org/abs/2207.04624 code：https://github.com/d1024choi/hlstrajforecast

半監(jiān)督學(xué)習(xí)/無監(jiān)督學(xué)習(xí)/自監(jiān)督學(xué)習(xí)

[1] FedX: Unsupervised Federated Learning with Cross Knowledge Distillation (FedX：具有交叉知識蒸餾的無監(jiān)督聯(lián)合學(xué)習(xí))paper：https://arxiv.org/abs/2207.09158

[2] Synergistic Self-supervised and Quantization Learning (協(xié)同自監(jiān)督和量化學(xué)習(xí))paper：https://arxiv.org/abs/2207.05432 code：https://github.com/megvii-research/ssql-eccv2022)

[3] Contrastive Deep Supervision (對比深度監(jiān)督)paper：https://arxiv.org/abs/2207.05306 code：https://github.com/archiplab-linfengzhang/contrastive-deep-supervision

[4] Dense Teacher: Dense Pseudo-Labels for Semi-supervised Object Detection (稠密教師：用于半監(jiān)督目標(biāo)檢測的稠密偽標(biāo)簽)paper：https://arxiv.org/abs/2207.02541

[5] Image Coding for Machines with Omnipotent Feature Learning (具有全能特征學(xué)習(xí)的機器的圖像編碼)paper：https://arxiv.org/abs/2207.01932

多模態(tài)學(xué)習(xí)/跨模態(tài)

視覺-語言

[1] Contrastive Vision-Language Pre-training with Limited Resources (資源有限的對比視覺語言預(yù)訓(xùn)練)paper：https://arxiv.org/abs/2112.09331code：https://github.com/zerovl/zerovl

跨模態(tài)

[1] Cross-modal Prototype Driven Network for Radiology Report Generation (用于放射學(xué)報告生成的跨模式原型驅(qū)動網(wǎng)絡(luò))paper：https://arxiv.org/abs/ code：https://github.com/markin-wang/xpronet

小樣本學(xué)習(xí)

[1] Learning Instance and Task-Aware Dynamic Kernels for Few Shot Learning (用于少數(shù)鏡頭學(xué)習(xí)的學(xué)習(xí)實例和任務(wù)感知動態(tài)內(nèi)核)paper：https://arxiv.org/abs/2112.03494

遷移學(xué)習(xí)/自適應(yīng)

[1] Factorizing Knowledge in Neural Networks (在神經(jīng)網(wǎng)絡(luò)中分解知識)paper：https://arxiv.org/abs/2207.03337 code：https://github.com/adamdad/knowledgefactor

[2] CycDA: Unsupervised Cycle Domain Adaptation from Image to Video (CycDA：從圖像到視頻的無監(jiān)督循環(huán)域自適應(yīng))paper：https://arxiv.org/abs/2203.16244

強化學(xué)習(xí)

[1] Target-absent Human Attention (目標(biāo)缺失——人類注意力缺失)paper：https://arxiv.org/abs/2207.01166 code：https://github.com/neouyghur/sess