RPA combines model-free and model-based RL, reaching 25.3% test success on R2R versus 23.1% for model-free RL.

Xin Wang, Wenhan Xiong, Hongmin Wang et al.

Proposes RotNet, using rotation recognition as self-supervised task to improve semantic feature learning, achieving near-supervised performance.

Spyros Gidaris, Praveer Singh, Nikos Komodakis

Proposes a category-specific deformable mesh model trained on image collections for single-view 3D reconstruction without ground-truth 3D data.

Angjoo Kanazawa, Shubham Tulsiani, Alexei A. Efros et al.

MAGSAC eliminates user-defined thresholds in RANSAC by marginalizing noise, significantly improving model accuracy.

Daniel Barath, Jana Noskova, Jiri Matas

Proposes an unsupervised monocular depth and visual odometry framework using deep feature reconstruction, outperforming prior methods on KITTI with a scale-consistent approach.

Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera et al.

LCR-Net++ integrates multi-scale pose proposals, classification, and regression for multi-person 2D/3D pose estimation, outperforming state-of-the-art methods.

Gregory Rogez, Philippe Weinzaepfel, Cordelia Schmid

Glimpse Clouds method uses visual attention for activity recognition without pose data, improving accuracy on NTU RGB+D dataset.

Fabien Baradel, Christian Wolf, Julien Mille et al.

Shield combines JPEG compression, vaccination, ensembling, and randomization to block 94% black-box and 98% gray-box attacks, enabling fast, practical defense.

Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen et al.

Transformers with local self-attention improve ImageNet negative log-likelihood from 3.83 to 3.77, surpassing PixelCNN++.

Niki Parmar, Ashish Vaswani, Jakob Uszkoreit et al.

Combining color and depth to improve SSCnet did not surpass the original method.

Andre Bernardes Soares Guedes, Teofilo Emidio de Campos, Adrian Hilton

Deep domain adaptation leverages adversarial, statistical, and reconstruction methods to improve cross-domain visual tasks, achieving up to 89.5% accuracy on benchmarks.

Mei Wang, Weihong Deng

DeepLabv3+ integrates ASPP and a simple decoder, using Xception backbone with depthwise separable convolution, achieving 89.0% mIOU on Pascal VOC 2012 without post-processing.

Liang-Chieh Chen, Yukun Zhu, George Papandreou et al.

RWalk combines EWC++ and Path Integral, using Riemannian geometry to improve incremental learning by balancing forgetting and intransigence.

Arslan Chaudhry, Puneet K. Dokania, Thalaiyasingam Ajanthan et al.

Proposed multi-task multi-modal CNN predicts steering and speed, using visual and feedback data, achieving high accuracy.

Zhengyuan Yang, Yixuan Zhang, Jerry Yu et al.

MobileNetV2 introduces inverted residuals and linear bottlenecks, achieving 72.0% Top-1 accuracy on ImageNet with only 3.4M parameters and 300M MAdd.

Mark Sandler, Andrew Howard, Menglong Zhu et al.

Deep MIL ranking-based weakly supervised anomaly detection achieves 0.85 AUC on a 128-hour surveillance dataset.

Waqas Sultani, Chen Chen, Mubarak Shah

Deep features as perceptual metrics outperform traditional methods, validated on large-scale human judgment datasets across architectures and supervision levels.

Richard Zhang, Phillip Isola, Alexei A. Efros et al.

Unified 3D deformation model captures face, hands, and body movements without markers, enabling multi-scale human motion tracking.

Hanbyul Joo, Tomas Simon, Yaser Sheikh

Proposes panoptic segmentation (PS) task combining semantic and instance segmentation, introducing the PQ metric for unified evaluation.

Alexander Kirillov, Kaiming He, Ross Girshick et al.

Proposes SuperPoint, a self-supervised interest point detector/describer using Homographic Adaptation, outperforming SIFT/ORB in real-time.

Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich