Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Threshold-Consistent Margin Loss for Open-World Deep Metric Learning

Authors: Qin ZHANG, Linghan Xu, Jun Fang, Qingming Tang, Ying Nian Wu, Joseph Tighe, Yifan Xing

ICLR 2024 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate TCM s effectiveness in enhancing threshold consistency while preserving accuracy, simplifying the threshold selection process in practical DML settings. 5 EXPERIMENTS
Researcher Affiliation Collaboration Qin Zhang1 , Linghan Xu1 , Qingming Tang2, Jun Fang1, Ying Nian Wu1, Joe Tighe1, Yifan Xing1 1 AWS AI Labs, 2 Alexa AI EMAIL, EMAIL
Pseudocode Yes Algorithm 1 Computation for OPIS metric and Algorithm 2 Training with TCM regularization
Open Source Code No The paper does not contain an explicit statement about releasing the source code for the described methodology or a link to a code repository.
Open Datasets Yes For training and evaluation, we use four commonly-used image retrieval benchmarks, namely i Naturalist-2018 (Horn et al., 2017), Stanford Online Product (Song et al., 2015), CUB-200-2011 (Wah et al., 2011) and Cars-196 (Krause et al., 2013).
Dataset Splits Yes The margin parameters (m+, mβˆ’) are tuned using grid search on 10% of the training data for each benchmark.
Hardware Specification Yes The Vi T-B/16 backbone is utilized with 8 Tesla V100 GPUs and a batch size of 392.
Software Dependencies No The paper mentions using the β€œtimm library (Wightman, 2019)” but does not specify its version number or the version numbers for other crucial software dependencies like PyTorch or Python.
Experiment Setup Yes During training, mini-batches are generated by randomly sampling 4 images per class following previous works (Brown et al., 2020; Patel et al., 2022). For TCM, we set Ξ»+ = Ξ»βˆ’ = 1. For OPIS, the calibration range is set to 1e-2 < FAR < 1e-1 for all benchmarks. The margin parameters (m+, mβˆ’) are tuned using grid search on 10% of the training data for each benchmark.