LSTD: A Low-Shot Transfer Detector for Object Detection

• Fully annotated target training sets may be limited, restricting the power of deep object detectors.Source: §1 · p. 1
• Transfer learning remains challenging because small target sets cannot readily eliminate detection–classification task differences and detectors overfit more readily than classifiers.Source: §1 · p. 1
• LSTD integrates SSD-style multi-convolutional bounding-box regression with Faster RCNN-style coarse-to-fine object classification.Source: §1 · p. 1, Figure 1 · p. 2
• TK transfers source-object-label knowledge to target proposals, while BD uses target bounding-box knowledge as extra feature-map supervision to suppress background disturbances.Source: §1 · p. 2
• LSTD outperforms other state-of-the-art approaches across challenging low-shot detection experiments.Source: §1 · p. 2

Object Detection

• Faster RCNN uses region proposals followed by Fast RCNN, whereas SSD is a one-stage detector with multi-layer bounding-box regression for varied object sizes.Source: §2 · p. 2
• Directly applying Faster RCNN or SSD with few examples is problematic because their architectures lack low-shot considerations.Source: §2 · p. 2

Low-shot Learning

• Existing low-shot learning approaches mainly address classification rather than object detection.Source: §2 · p. 2

LSTD

• LSTD combines Faster RCNN and SSD advantages and uses regularization to integrate object knowledge from source and target domains.Source: §2 · pp. 2–3

Basic Deep Architecture of LSTD

• Pretrained SSD regression parameters can initialize the target domain, avoiding random re-initialization of bounding-box regression with few target images.This reduces fine-tuning burdens when target data are scarce.Source: §3.1 · p. 3
• LSTD combines SSD-style multi-layer bounding-box regression with Faster RCNN-style coarse-to-fine object classification in a unified architecture.The design separates regression and classification across relatively separate places to decompose low-shot learning difficulties.Source: §3.1 · pp. 3–4

Regularized Transfer Learning for LSTD

• LSTD fine-tunes a source-domain detector on the target domain with a regularization combining background-depression and transfer-knowledge terms.The total regularization uses coefficients λBD and λTK for LBD and LTK.Source: §3.2 · p. 4
• BD masks middle-level feature cubes using target ground-truth boxes and penalizes background activation, helping suppress background regions and focus on target objects.Figure 3 reports reduced background disturbances on feature heatmaps.Source: §3.2 · p. 4, Figure 3 · p. 4
• TK supplies source-object knowledge for target proposals despite different categories, using source-class soften-labels to transfer relations such as color or shape similarity.Figure 4 illustrates relevant source categories for Cow/Aeroplane proposals through Bear/Kite soften-labels.Source: §3.2 · p. 4, Figure 4 · p. 4

Pre-training on Large-scale Source Domain

• Source-domain LSTD is trained with a large-scale detection benchmark.Source: §1 · p. 5

Initialization for Small-scale Target Domain

• The last layer of the (K + 1)-object classifier is randomly initialized because source and target object categories differ.All other parts of target-domain LSTD are initialized from source-domain LSTD.Source: §2 · p. 5

Fine-tuning for Small-scale Target Domain

• Target-domain LSTD is fine-tuned with BD and TK regularizations using the small training set.Source: §3 · p. 5
• Target-domain proposals are passed through the source-domain LSTD ROI pooling and object classifier to generate source-domain knowledge vectors.Source: §3 · p. 5
• The source classifier converts proposal activations into soften-labels using softmax with temperature τ > 1, retaining richer label-relation information.Source: §3 · p. 5
• A source-object soften classifier is added to target-domain LSTD so each target proposal predicts source-object categories as an auxiliary task.Source: §3 · p. 5
• Cross-entropy between source-domain knowledge and target predictions forms the TK regularization, integrating source knowledge into target training.Source: §3 · p. 5

Experimental setup

• LSTD is evaluated on three transfer tasks using large-scale source training and 1/2/5/10/30 target training images per object class.The datasets include COCO, ImageNet2015, VOC2007, and VOC2010.Source: §4 · p. 5

Basic Deep Structure of LSTD

• In Task 1, LSTD achieves competitive source-domain mAP and significantly outperforms SSD and Faster RCNN for one-shot target detection.The comparison uses models fine-tuned from source-domain pretrained models.Source: §4.1 · p. 6

Regularized Transfer Learning for LSTD

• TK and BD regularizations significantly improve LSTD fine-tuning over the baseline, especially when the target training set is scarce, such as in one-shot detection.BD remains generally robust when implemented on different convolutional layers.Source: Table 3 · p. 6, §4.1 · p. 7

Comparison with the State-of-the-art

• LSTD outperforms SSD and Faster RCNN in Task 1 as target-shot counts change, and surpasses weakly- or semi-supervised detectors beyond two shots in Tasks 2 and 3.The weakly- and semi-supervised results are taken from their original papers on the same test set.Source: Figure 5 · p. 6, §4.2 · p. 7
• LSTD outperforms competing detectors on Task 3 with 0.4% of training data and is competitive with fully supervised detectors using 11% of the training set.At 11%, the reported mAP values are LSTD:69.7, SSD:68.0, and Faster RCNN:69.9.Source: §4.2 · p. 7

Visualization

• Visualization shows that one-shot LSTD can localize but sometimes misclassifies objects, while five-shot training largely clarifies these errors.For 2-shot LSTD on VOC 2007, Animal errors are reported as 23% localization, 71% similarity, and 4% background.Source: §4.3 · p. 7, Figure 6 · p. 7

• LSTD addresses object detection with a few training images through a flexible deep architecture designed to reduce low-shot transfer difficulties.Source: §5 · p. 7
• The architecture incorporates strengths of SSD and Faster RCNN within a unified deep framework.Source: §5 · p. 7
• TK and BD terms form a regularized transfer-learning framework intended to generalize fine-tuning with few target images.Source: §5 · p. 7
• LSTD outperforms other state-of-the-art approaches across challenging low-shot detection experiments.Source: §5 · p. 7