Source-linked AI summary
Learning Deep Feature Representations with Domain Guided Dropout for Person Re-identification
Tong Xiao, Hongsheng Li, Wanli Ouyang, Xiaogang Wang
TL;DR
The paper asks how to learn generic, robust person re-identification features from multiple biased domains when no single dataset is large enough. It jointly trains a CNN across domains and introduces Domain Guided Dropout to adapt neuron regularization by domain. The resulting pipeline and method outperform prior methods across multiple person re-identification datasets, with a reported largest gain of 46% on PRID.
Problem
Multiple datasets for the same problem may be too small individually and biased toward limited variations, creating a need to jointly learn generic feature representations.
Method
The paper mixes multiple domains to train a person-identity CNN, then uses Domain Guided Dropout to assign domain-specific dropout rates to neurons based on their effectiveness.
Results
46% was the largest reported gain on PRID, and the methods outperformed state-of-the-art approaches by large margins across multiple person re-identification datasets.
Takeaways & Limitations
Jointly using multiple datasets and domain-guided neuron regularization improves feature learning for person re-identification across domains.
Abstract
from arXiv · showhide
Learning generic and robust feature representations with data from multiple domains for the same problem is of great value, especially for the problems that have multiple datasets but none of them are large enough to provide abundant data variations. In this work, we present a pipeline for learning deep feature representations from multiple domains with Convolutional Neural Networks (CNNs). When training a CNN with data from all the domains, some neurons learn representations shared across several domains, while some others are effective only for a specific one. Based on this important observation, we propose a Domain Guided Dropout algorithm to improve the feature learning procedure. Experiments show the effectiveness of our pipeline and the proposed algorithm. Our methods on the person re-identification problem outperform state-of-the-art methods on multiple datasets by large margins.
1. Introduction
The paper addresses learning generic person re-identification features from multiple biased domains when no single dataset provides enough variation. It combines joint CNN training with Domain Guided Dropout, which adapts neuron dropout to domain effectiveness.
- Motivation: Multiple smaller datasets often target similar problems but differ in data distributions, motivating algorithms that jointly use them to learn generic representations.Multi-domain learning uses data across domains simultaneously, while many specific areas lack a large-scale dataset for robust feature learning.
- Motivation: Person re-identification datasets capture different scene and imaging conditions, so each dataset contains only a subset of possible variations.Examples include campus imagery with backpacks, street views with crosswalks, and substantial cross-camera resolution changes.
- Approach: The proposed pipeline mixes domains and trains a CNN to recognize person identities, targeting representations effective across all domains simultaneously.The paper uses a CNN whose capacity is designed for the scale of the mixed dataset.
- Approach: Domain Guided Dropout assigns each neuron a domain-specific dropout rate to mute neurons considered unrelated to that domain.It differs from standard Dropout, which treats all neurons equally, and includes deterministic and stochastic schemes.
- Results: 46% was the largest reported gain from using multiple datasets, occurring on the PRID dataset, while the method outperformed state-of-the-art methods by large margins on multiple datasets.The paper attributes these results to jointly learned features and domain-specific neuron regularization.
2. Related Work
The related work situates the paper within multi-domain representation learning, neural-network regularization, and person re-identification methods. It contrasts Domain Guided Dropout with existing distribution-alignment and adaptive-dropout approaches while noting complementary metric-learning work.
- Multi-domain learning: Prior multi-domain learning work includes neural-network regularization with multiple-kernel MMD and methods that reduce source-target distribution mismatch.These approaches address differences between domains through distribution-based objectives or adaptation.
- Neural-network regularization: Dropout is a widely used neural-network regularizer whose quality affects the discriminative power and generalization ability of trained models.The related work frames regularization as a way to reduce the effective parameter space of networks with millions of parameters.
- Neural-network regularization: Adaptive dropout predicts neuron-specific dropout probabilities from neuron responses, whereas this paper uses domain information to guide dropout during multi-domain CNN training.The distinction is that the proposed approach conditions dropout on domain effectiveness rather than only on individual neuron responses.
- Person re-identification: Person re-identification research addresses feature representation and metric learning, including methods based on handcrafted features, triplet samples, and learned distances.Metric-learning methods provide a complementary route to improving identity matching beyond feature learning.
3. Method
The method jointly trains a CNN on multiple person re-identification domains, measures neuron effectiveness per domain, and uses those scores to guide domain-specific dropout during continued training.
- Joint learning objective and CNN structure: The pipeline mixes data and labels from all domains, then trains a CNN with a single softmax loss to produce a shared baseline.The joint model is trained from scratch on the merged dataset before domain-specific neuron analysis.
- Problem formulation: The learned feature extractor is evaluated by comparing Euclidean distances between probe and gallery features during person re-identification.The goal is similar features for images of the same person and dissimilar features for different people.
- Domain Guided Dropout: The method estimates each neuron’s domain-specific impact as the loss increase caused by setting that neuron’s response to zero, then averages impacts over domain samples.The exact computation requires many forward passes, so the method also studies a Taylor-based approximation.
- Domain Guided Dropout: Different domains have different effective neurons: paired-domain impact-score curves show little correlation after neurons are sorted by one domain’s scores.This supports treating neuron usefulness as domain dependent rather than uniform across the joint dataset.
- Domain Guided Dropout: Domain Guided Dropout uses impact scores to construct domain-specific masks that mute nonrelated neurons during continued CNN training.The deterministic scheme discards neurons with nonpositive scores, while the stochastic scheme samples masks from score-dependent probabilities.
- Domain Guided Dropout: The stochastic scheme approaches deterministic masking as T → 0 and standard Dropout with ratio 0.5 as T → ∞.The temperature controls how strongly impact scores affect retention probabilities.
4. Experiments
Experiments evaluate multi-domain person re-identification across varied datasets and compare joint learning, Domain Guided Dropout, fine-tuning, and prior methods. Results show that joint training and domain-specific dropout improve performance, with deterministic and stochastic variants best suited to different training settings.
- Datasets and protocols: Experiments use seven person re-identification datasets spanning varied scales, scenes, viewpoints, lighting conditions, and occlusions, with six datasets used for testing.Shinpuhkan is used only for training because it contains 24 people; the other datasets follow established probe and gallery protocols.
- Comparison methods: The evaluation compares individual-domain training, joint single-task learning, deterministic Domain Guided Dropout, and stochastic Domain Guided Dropout fine-tuning.The joint model provides one model for all domains, while stochastic dropout is additionally used for domain-specific fine-tuning.
- Comparison with state-of-the-art methods: 72.6% CMC top-1 accuracy on CUHK03 exceeds the previous best deep learning result of 54.7% by 18%.The method uses a single CNN trained with identity classification and Euclidean distance rather than a two-stream verification network.
- Joint learning: Joint learning improves performance on most datasets, while CUHK03 decreases slightly when domains are combined without special treatment.The authors hypothesize that larger domains help others but may reduce discriminability on themselves; balancing sources in minibatches can cause severe overfitting.
- Domain Guided Dropout: 0.5%-2.7% gains across all domains show that deterministic Domain Guided Dropout improves the converged joint model through domain-specific regularization.The method computes neuron impact scores by domain and replaces standard dropout in the pretrained JSTL model.
- Domain Guided Dropout: Fine-tuning with stochastic Domain Guided Dropout increases final CMC top-1 accuracies by several percent and improves over standard-dropout fine-tuning by 3% on average.This adaptation targets domain-specific biases but sacrifices generalization to other domains; deterministic dropout is better for joint training, whereas stochastic dropout is better for separate fine-tuning.
5. Conclusion
The paper presents a pipeline for learning generic and robust CNN feature representations from multiple domains and proposes Domain Guided Dropout to improve feature learning. Experiments on multiple person re-identification datasets report large-margin improvements over state-of-the-art methods.
- The paper addresses learning generic and robust CNN feature representations from multiple domains.
- The proposed pipeline and Domain Guided Dropout algorithm are evaluated through extensive experiments on multiple person re-identification datasets.
- The methods outperform state-of-the-art approaches by large margins on most evaluated datasets.