Instant-Fold enables execution of multiple deformable object manipulation modes from a single demonstration via a flow-matching transformer policy that transfers zero-shot from simulation to real robots.
hub
In Defense of the Triplet Loss for Person Re-Identification
38 Pith papers cite this work. Polarity classification is still indexing.
38
Pith papers citing it
abstract
In the past few years, the field of computer vision has gone through a revolution fueled mainly by the advent of large datasets and the adoption of deep convolutional neural networks for end-to-end learning. The person re-identification subfield is no exception to this. Unfortunately, a prevailing belief in the community seems to be that the triplet loss is inferior to using surrogate losses (classification, verification) followed by a separate metric learning step. We show that, for models trained from scratch as well as pretrained ones, using a variant of the triplet loss to perform end-to-end deep metric learning outperforms most other published methods by a large margin.
hub tools
citation-role summary
method 1
citation-polarity summary
roles
method 1polarities
use method 1representative citing papers
ToolFG is the first tool-integrated MLLM framework for fine-grained image classification, trained via MCTS-guided distillation from proprietary models and refined through model-tool co-evolution.
Standard vehicle re-ID benchmarks allow memorization of seen vehicle types; a new train/test split by vehicle type and view shows that state-of-the-art methods fail to generalize to unseen vehicles.
CityGuard introduces a graph-aware transformer with dispersion-adaptive metrics, spatially conditioned attention using coarse geometry, and differentially private embeddings to support privacy-preserving identity retrieval across city cameras.
RAG-HAR combines retrieval-augmented generation with LLMs to deliver state-of-the-art human activity recognition across six benchmarks without any model training or fine-tuning.
SAGA-ReID improves CLIP-based person ReID by using structured anchor-guided aggregation of patch tokens, delivering up to 10.6 Rank-1 gains on occluded benchmarks over global pooling.
GoD uses anatomy graphs and difference alignment to improve medical image re-identification accuracy and auditability, with +7.1 pp Rank-1 gains on fundus and +3.1 pp on CXR.
TimeProVe proposes a propose-then-verify framework using lightweight action-based candidate evidence generation followed by targeted VLM verification for efficient long video temporal reasoning, achieving 7.3% improvement on OTB with 75% fewer VLM calls.
Ortho-ReID learns instance-adaptive low-rank orthogonal subspaces from VLM text via a transformer Basis Maker and enforces geometric constraints to produce clothing-invariant identity features, reporting SOTA gains on PRCC, Celeb-reID-light, and LaST.
MERIT trains disentangled heads for melody, rhythm, and timbre via conditional audio generation and stem separation, with evaluations showing each head responds strongly to its target dimension and near chance on others across synthetic and real audio.
DMDL debias modality cues at model and optimization levels via causal adjustment intervention and collaborative bias-free training to learn modality-invariant features for unsupervised VI-ReID.
SD-ReID trains a ViT to extract identity and view conditions, fine-tunes Stable Diffusion to generate view-mimicking features, adds a View-Refined Decoder, and combines both identity and all-view features for retrieval on aerial-ground re-identification benchmarks.
LVLM-ReID guides LVLMs to produce refined semantic tokens as pedestrian identity features for ReID, achieving competitive benchmark results without additional image-text data.
STCnet recovers occluded parts in video person re-ID using spatio-temporal cues to form the VRSTC framework, outperforming prior methods on three datasets.
A teacher-student model with co-saliency network and growing-probability occlusion simulator outperforms prior methods on four occluded person re-identification benchmarks.
Neural features from Twitch streams are pooled via hierarchical Bayesian model to estimate CS:GO gamer intrinsic skill, validated by correlation with subsequent public ranks.
MELD is a multi-task AI-text detector using auxiliary heads, uncertainty-weighted losses, EMA distillation, and pairwise ranking that reaches 99.9% TPR at 1% FPR on a new held-out benchmark while remaining competitive on the RAID leaderboard.
PAD uses prompt distillation on the text side and domain-adaptive EMA prompts on the visual side to balance stability and plasticity in lifelong person re-identification.
A new benchmark dataset and competition for top-view RGB-Depth person re-identification is released, with competition results showing RGB easier than depth and cross-modal retrieval.
Self-supervised ReLU networks form substantially fewer linear regions than supervised models for comparable accuracy, with contrastive methods rapidly expanding regions and self-distillation consolidating them, enabling early geometric detection of representation collapse.
ReID-R achieves competitive person re-identification performance using chain-of-thought reasoning and reinforcement learning with only 14.3K non-trivial samples, about 20.9% of typical data scales, while providing interpretations.
CraterBench-R is a new retrieval benchmark where self-supervised ViTs with a training-free instance-token aggregation method achieve high accuracy for identifying individual craters while reducing storage needs.
DNSMOS-C adds MOS-guided triplet contrastive loss to DNSMOS Pro for improved correlation, out-of-domain generalization, and emergent low-dimensional quality ordering in embeddings via unified training.
ConcernBERT is a BERT embedding model trained with triplet loss on class membership to encode concern-level semantics in Java entities, evaluated by recovering original classes from merged unlabeled groups on a new dataset of over 2M files, outperforming existing models.
citing papers explorer
-
Instant-Fold: In-Context Imitation Learning for Deformable Object Manipulation
Instant-Fold enables execution of multiple deformable object manipulation modes from a single demonstration via a flow-matching transformer policy that transfers zero-shot from simulation to real robots.
-
ToolFG: Towards Well-Grounded Fine-Grained Image Classification
ToolFG is the first tool-integrated MLLM framework for fine-grained image classification, trained via MCTS-guided distillation from proprietary models and refined through model-tool co-evolution.
-
Generalization Limits in Vehicle Re-Identification
Standard vehicle re-ID benchmarks allow memorization of seen vehicle types; a new train/test split by vehicle type and view shows that state-of-the-art methods fail to generalize to unseen vehicles.
-
CityGuard: Graph-Aware Private Descriptors for Bias-Resilient Identity Search Across Urban Cameras
CityGuard introduces a graph-aware transformer with dispersion-adaptive metrics, spatially conditioned attention using coarse geometry, and differentially private embeddings to support privacy-preserving identity retrieval across city cameras.
-
RAG-HAR: Retrieval Augmented Generation-based Human Activity Recognition
RAG-HAR combines retrieval-augmented generation with LLMs to deliver state-of-the-art human activity recognition across six benchmarks without any model training or fine-tuning.
-
From Global to Local: Rethinking CLIP Feature Aggregation for Person Re-Identification
SAGA-ReID improves CLIP-based person ReID by using structured anchor-guided aggregation of patch tokens, delivering up to 10.6 Rank-1 gains on occluded benchmarks over global pooling.
-
Graph-of-Differences: Anatomy-Structured Difference Alignment for Medical Image Re-Identification
GoD uses anatomy graphs and difference alignment to improve medical image re-identification accuracy and auditability, with +7.1 pp Rank-1 gains on fundus and +3.1 pp on CXR.
-
TimeProVe: Propose, then Verify for Efficient Long Video Temporal Reasoning in Activities of Daily Living
TimeProVe proposes a propose-then-verify framework using lightweight action-based candidate evidence generation followed by targeted VLM verification for efficient long video temporal reasoning, achieving 7.3% improvement on OTB with 75% fewer VLM calls.
-
Learning Instance-Adaptive Low-Rank Orthogonal Subspaces for Clothes-Changing Person Re-Identification
Ortho-ReID learns instance-adaptive low-rank orthogonal subspaces from VLM text via a transformer Basis Maker and enforces geometric constraints to produce clothing-invariant identity features, reporting SOTA gains on PRCC, Celeb-reID-light, and LaST.
-
MERIT: Learning Disentangled Music Representations for Audio Similarity
MERIT trains disentangled heads for melody, rhythm, and timbre via conditional audio generation and stem separation, with evaluations showing each head responds strongly to its target dimension and near chance on others across synthetic and real audio.
-
Dual-level Modality Debiasing Learning for Unsupervised Visible-Infrared Person Re-Identification
DMDL debias modality cues at model and optimization levels via causal adjustment intervention and collaborative bias-free training to learn modality-invariant features for unsupervised VI-ReID.
-
SD-ReID: View-aware Stable Diffusion for Aerial-Ground Person Re-Identification
SD-ReID trains a ViT to extract identity and view conditions, fine-tunes Stable Diffusion to generate view-mimicking features, adds a View-Refined Decoder, and combines both identity and all-view features for retrieval on aerial-ground re-identification benchmarks.
-
When Large Vision-Language Models Meet Person Re-Identification
LVLM-ReID guides LVLMs to produce refined semantic tokens as pedestrian identity features for ReID, achieving competitive benchmark results without additional image-text data.
-
VRSTC: Occlusion-Free Video Person Re-Identification
STCnet recovers occluded parts in video person re-ID using spatio-temporal cues to form the VRSTC framework, outperforming prior methods on three datasets.
-
A Novel Teacher-Student Learning Framework For Occluded Person Re-Identification
A teacher-student model with co-saliency network and growing-probability occlusion simulator outperforms prior methods on four occluded person re-identification benchmarks.
-
E-Sports Talent Scouting Based on Multimodal Twitch Stream Data
Neural features from Twitch streams are pooled via hierarchical Bayesian model to estimate CS:GO gamer intrinsic skill, validated by correlation with subsequent public ranks.
-
MELD: Multi-Task Equilibrated Learning Detector for AI-Generated Text
MELD is a multi-task AI-text detector using auxiliary heads, uncertainty-weighted losses, EMA distillation, and pairwise ranking that reaches 99.9% TPR at 1% FPR on a new held-out benchmark while remaining competitive on the RAID leaderboard.
-
Prompt-Anchored Vision-Text Distillation for Lifelong Person Re-identification
PAD uses prompt distillation on the text side and domain-adaptive EMA prompts on the visual side to balance stability and plasticity in lifelong person re-identification.
-
ICPR 2026 Competition on Privacy-Preserving Person Re-Identification from Top-View RGB-Depth Camera (TVRID)
A new benchmark dataset and competition for top-view RGB-Depth person re-identification is released, with competition results showing RGB easier than depth and cross-modal retrieval.
-
Complexity of Linear Regions in Self-supervised Deep ReLU Networks
Self-supervised ReLU networks form substantially fewer linear regions than supervised models for comparable accuracy, with contrastive methods rapidly expanding regions and self-distillation consolidating them, enabling early geometric detection of representation collapse.
-
Thinking Before Matching: A Reinforcement Reasoning Paradigm Towards General Person Re-Identification
ReID-R achieves competitive person re-identification performance using chain-of-thought reasoning and reinforcement learning with only 14.3K non-trivial samples, about 20.9% of typical data scales, while providing interpretations.
-
CraterBench-R: Instance-Level Crater Retrieval for Planetary Scale
CraterBench-R is a new retrieval benchmark where self-supervised ViTs with a training-free instance-token aggregation method achieve high accuracy for identifying individual craters while reducing storage needs.
-
DNSMOS-C: Improving End-to-end Speech Quality Models via Contrastive Learning
DNSMOS-C adds MOS-guided triplet contrastive loss to DNSMOS Pro for improved correlation, out-of-domain generalization, and emergent low-dimensional quality ordering in embeddings via unified training.
-
ConcernBERT: Learning Responsibilities Using Class Membership
ConcernBERT is a BERT embedding model trained with triplet loss on class membership to encode concern-level semantics in Java entities, evaluated by recovering original classes from merged unlabeled groups on a new dataset of over 2M files, outperforming existing models.
-
Aligning Implied Statements for Implicit Hate Speech Generalizability with Context-Bounded Semi-hard Negative Mining
ImpSH improves cross-domain generalization in implicit hate speech classification by aligning posts with implied statements and applying context-bounded semi-hard negative mining within a triplet learning setup.
-
HOTFLoc++: End-to-End Hierarchical LiDAR Place Recognition, Re-Ranking, and 6-DoF Metric Localisation in Forests
A hierarchical octree-transformer framework for LiDAR place recognition, multi-scale re-ranking, and 6-DoF localization in forests reports 90.7% Recall@1 on CS-Wild-Places and under 2 m / 5° error on 97.2% of registrations with nearly 100x runtime gains over RANSAC.
-
A Novel Method for News Article Event-Based Embedding
Proposes an event-based news embedding method via entity/theme extraction, periodic GloVe models, SIF, and Siamese networks, claiming outperformance on shared event detection using GDELT data.
-
Hard-Aware Fashion Attribute Classification
Presents HABP to emphasize hard samples during training and Deact to generate stable synthetic samples for rare attributes, outperforming prior methods on large-scale fashion datasets without extra supervision.
-
Beyond Pedestrians: Caption-Guided CLIP Framework for High-Difficulty Video-based Person Re-Identification
CG-CLIP adds caption-guided memory refinement and token-based spatiotemporal aggregation to CLIP for video person ReID, outperforming SOTA on MARS, iLIDS-VID, SportsVReID and DanceVReID.
-
Spatial-Temporal Expert Learning for Video-based Person Re-identification
Proposes dynamic expert selection with input-aware and spatial-temporal mechanisms plus an extendable scheme to improve fine-grained feature use in video person Re-ID.
-
Privacy-Preserving Person Re-Identification from Temporal Sequences with Transformer and Hungarian Optimization
Depth-only Re-ID via Transformer on temporal sequences and Hungarian matching achieves competitive accuracy on top-view datasets while prioritizing privacy.
-
HDST-GNN: Heterogeneous Dynamic Spatiotemporal Graph Neural Networks for Multi-Object Tracking in UAV Aerial Imagery
The paper introduces HDST-GNN, a heterogeneous dynamic spatiotemporal GNN for UAV multi-object tracking with altitude-adaptive edges, typed nodes, and occlusion-gated aggregation, reporting 94.51% MOTA on VisDrone2019-MOT.
-
Knowing When Not to Predict: Self Supervised Learning and Abstention for Safer DR Screening
SSL pretraining enhances calibrated confidence and selective performance in DR screening, yet benefits on reliability plateau and longer pretraining does not reliably improve abstention.
-
Few-Shot Network Intrusion Detection Using Online Triplet Mining
A triplet network using online triplet mining and KNN classifier achieves competitive few-shot performance on network intrusion detection with as few as 10 malicious samples per class.
-
Enhancing the Discriminative Feature Learning for Visible-Thermal Cross-Modality Person Re-Identification
EDFL improves visible-thermal person re-identification by enhancing feature discriminability with skip connections and dual-modality triplet loss, outperforming state-of-the-art on two datasets.
-
ThirdEye: Triplet Based Iris Recognition without Normalization
ThirdEye applies triplet convolutional neural networks directly to segmented iris images without normalization, reporting EERs of 1.32% on ND-0405, 9.20% on UbirisV2, and 0.59% on IITD, improving prior results on the constrained IITD set.
-
On the Properties of Feature Attribution for Supervised Contrastive Learning
Neural networks trained via supervised contrastive learning yield feature attributions that are more faithful, less complex, and more continuous than those from cross-entropy trained networks.
-
Identity-Aware U-Net: Fine-grained Cell Segmentation via Identity-Aware Representation Learning
Identity-Aware U-Net augments a U-Net backbone with an auxiliary embedding branch and triplet metric learning to discriminate among cells with near-identical shapes and textures.