Weakly Supervised Basic Detection Network

• The paper demonstrates that joint optimization of MIL-based proposal mining, box regression, and guided attention significantly improves weakly supervised detection performance.
• WSBDN leverages a unified architecture with a VGG16 backbone, selective search proposals, and attention mechanisms to enhance localization accuracy.
• Empirical results on PASCAL VOC benchmarks reveal notable mAP gains, confirming WSBDN's robustness over traditional two-phase weakly supervised detection pipelines.

Within weakly supervised object detection (WSOD), a Weakly Supervised Basic Detection Network (WSBDN) denotes a weakly supervised version of a basic detector: a shared feature extractor followed by proposal-level classification and bounding-box regression, with proposal-level supervision bootstrapped from a multiple instance learning (MIL) branch rather than supplied by human bounding boxes. In "Towards Precise End-to-end Weakly Supervised Object Detection Network," the term is used conceptually for a single end-to-end network that couples MIL-based proposal mining, a Fast R-CNN-like regression branch, and a guided attention module (GAM) under image-level supervision only (Yang et al., 2019). The designation is not universal across the WSOD literature: "Collaborative Learning for Weakly Supervised Object Detection" uses a WSDDN-like weak detector inside a collaborative framework but does not introduce or evaluate a method named WSBDN (Wang et al., 2018).

1. Concept and nomenclature

WSBDN, as described in the 2019 end-to-end WSOD formulation, is a weakly supervised analogue of a basic detector in which the same proposal features support both recognition and localization. The decisive distinction from a fully supervised detector is that proposal-level targets are not available directly from bounding-box annotation; instead, they are inferred online from image-level labels through a MIL branch. This makes the WSBDN formulation structurally close to standard region-based detectors while preserving the annotation regime of weak supervision (Yang et al., 2019).

The conceptual role of WSBDN is therefore not to replace MIL, but to embed MIL into a detector that also performs box regression. The MIL branch produces pseudo ground-truth proposals, and the detection branch classifies each RoI and regresses its box. The architecture is designed to avoid the common WSOD pattern in which proposal mining and box refinement are separated into two phases. A plausible implication is that the term "WSBDN" is best understood not as a universally standardized model name, but as a description of a particular weakly supervised detector design in which proposal classification and regression are learned jointly.

2. Architectural composition

The end-to-end WSBDN-style network has three principal components: a shared CNN backbone, a MIL-based detection branch, and a bounding-box regression / detection branch. In the reported experiments, the backbone is VGG16 pretrained on ImageNet. Region proposals are generated using Selective Search, and RoI pooling extracts fixed-size region features from the backbone feature maps for downstream processing (Yang et al., 2019).

The MIL branch is based on the WSDDN / OICR family. It takes RoI features, computes proposal-level class scores, aggregates them into image-level predictions, and is trained using only image-level labels. Its explicit function is proposal discovery: it supplies pseudo ground-truth boxes for subsequent refinement. The regression branch is Fast R-CNN-like and uses the same RoI features to train a detection head with a classification sibling and a bounding-box regression sibling. Because the two branches share the backbone, the discovery and refinement processes remain tightly coupled throughout training.

This arrangement is the "basic detection" aspect of WSBDN. The network performs proposal-level classification and box regression as in a conventional detector, but the supervision path is indirect. Rather than starting from box labels, the detector starts from pseudo boxes mined by MIL and updates them online during optimization.

3. Guided attention and implicit localization

A central addition to the WSBDN formulation is the guided attention module. The motivating observation is that a well trained classification network contains rich object location information, even when only image-level labels are provided. GAM is added on top of the backbone so that the shared features become more localization-aware before RoI pooling and detection (Yang et al., 2019).

The module begins from a conventional attention mechanism and produces an attended feature map $\hat{X}$. The residual form of the attention update is emphasized: the feature is enhanced rather than replaced, so the attended representation preserves the original signal while amplifying task-relevant responses. The formulation is then extended from purely spatial attention to spatial + channel attention, yielding channel-specific modulation of the feature tensor.

The "guided" aspect arises from supervision. The attention map is passed through another convolution and a global average pooling layer to produce an image-level class score vector, and a standard multi-label classification loss is applied. In consequence, the attention mechanism is trained to highlight regions that improve class prediction. The paper interprets this as implicit location supervision: foreground regions become salient because they are the regions most useful for solving the image-label task.

4. Optimization objective and learning signals

The optimization combines image-level classification, proposal mining, classifier refinement, and detector training in a single end-to-end objective. For image-level labels $Y=[y_1,\dots,y_C]$, the WSDDN-style MIL branch produces image-level class probabilities $p_c$, and the MIL loss is

$$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$

Once pseudo ground-truth boxes are generated, the detection branch uses a Fast R-CNN-style multi-task loss,

$$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$

where $\mathcal{L}_{cls}$ is RoI classification loss, $\mathcal{L}_{loc}$ is box regression loss using smooth $L_1$, and $\lambda$ balances classification and localization. Because pseudo ground truth is noisy, the classification term is confidence-weighted per proposal:

$$\mathcal{L}_{cls} = - \frac{1}{|R|}\sum_{r=1}^{|R|}\sum_{c=1}^{C+1} w^r u_c^r \log p_c^r.$$

The full optimization is the sum of four components: $Y=[y_1,\dots,y_C]$0 for GAM, $Y=[y_1,\dots,y_C]$1 for proposal mining, $Y=[y_1,\dots,y_C]$2 for classifier refinement from OICR/PCL, and $Y=[y_1,\dots,y_C]$3 for the detection branch. This formulation makes the system end-to-end in a literal sense: attention learning, MIL discovery, refinement, and box regression are optimized together rather than staged sequentially (Yang et al., 2019).

5. Local minima in two-phase WSOD

The principal rationale for WSBDN is the failure mode of two-phase WSOD. In a common pipeline, MIL first discovers proposals based on discriminative evidence, and a second-stage detector is then trained from those pseudo labels. The difficulty is that MIL often locks onto parts rather than full objects. For categories such as cat, the discovered proposal may be the head instead of the body. If such part boxes are treated as pseudo ground truth for a second-stage regressor, the detector can refine the wrong region accurately and become trapped in a local minimum (Yang et al., 2019).

The joint WSBDN strategy changes the timing of refinement. The regressor begins learning while the MIL branch is still relatively diffuse and before it has fully collapsed onto small discriminative parts. The paper argues that this allows the regression branch to counteract part-based bias early in training. The qualitative comparison reported in the paper supports this interpretation: the MIL detector alone shifts toward object parts over training, whereas the joint model maintains fuller object localization.

This suggests that the principal contribution of WSBDN is not merely the presence of regression, but the synchronization of regression with weak proposal discovery. Localization refinement starts early enough to influence which proposals become dominant.

6. Empirical characteristics and implementation

The method is evaluated on PASCAL VOC 2007 and 2012 with mAP on test sets and CorLoc on trainval. The reported results place the model among state-of-the-art weakly supervised detectors in both single-model and ensemble settings. On VOC 2007 test, the single-model OICR-based version with GAM+REG reaches $Y=[y_1,\dots,y_C]$4 mAP, compared with an OICR baseline at $Y=[y_1,\dots,y_C]$5 mAP. On VOC 2012 test, the single-model version reaches $Y=[y_1,\dots,y_C]$6 mAP or $Y=[y_1,\dots,y_C]$7 mAP depending on the MIL backend. Ensemble variants reach $Y=[y_1,\dots,y_C]$8 mAP on VOC 2007 and $Y=[y_1,\dots,y_C]$9 CorLoc on VOC 2007 trainval (Yang et al., 2019).

The ablation results isolate the effects of guided attention and joint regression. On VOC 2007 test, the MIL baseline reaches $p_c$0 mAP; MIL + GAM reaches $p_c$1 mAP; MIL + FRCN reaches $p_c$2 mAP; MIL + REG reaches $p_c$3 mAP; and MIL + GAM + REG reaches $p_c$4 mAP. The corresponding CorLoc values are $p_c$5, $p_c$6, $p_c$7, and $p_c$8 for the principal variants reported. The largest gains are observed for categories such as cat and dog, which are explicitly identified as suffering from part-based collapse.

Implementation details are also reported with some precision. The method uses VGG16 pretrained on ImageNet, Selective Search windows, and Caffe. Training uses a mini-batch size of $p_c$9 images, a learning rate of $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$0 for $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$1K iterations followed by $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$2 for $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$3K iterations, momentum $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$4, and weight decay $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$5. New layers are initialized from $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$6 and use a $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$7 learning rate. Multi-scale training and testing employ image scales $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$8, horizontal flips are used during training and testing, and evaluation adopts an IoU threshold of $$\mathcal{L}_{mil} = - \sum_{c=1}^{C}\{y_c\log p_c \ + \ (1-y_c)\log(1-p_c)\}.$$9 (Yang et al., 2019).

7. Relation to WSDDN-like detectors and collaborative learning

WSBDN should not be conflated with WSDDN, although the two are closely related. In the collaborative framework of "Collaborative Learning for Weakly Supervised Object Detection," the weakly supervised branch $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$0 is explicitly a WSDDN-like two-stream proposal detector. One stream computes classification scores $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$1, another computes localization scores $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$2, the proposal-level detection score is $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$3, and image-level prediction is formed by aggregation, $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$4. The weak branch is trained with image-level multi-label binary cross-entropy, while a Faster R-CNN-like strong branch $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$5 is trained without box labels through prediction consistency and partial feature sharing (Wang et al., 2018).

That framework is directly relevant to WSBDN because it shows how a WSDDN-style weak detector can be coupled to a region-based detector without strong supervision. However, it is not a WSBDN paper in the terminological sense: the paper repeatedly refers to a WSDDN-like architecture, not to WSBDN. Its contribution is Weakly Supervised Collaborative Learning (WSCL), instantiated as a Weakly Supervised Collaborative Detection Network (WSCDN), with shared convolutional layers and shared bottom fully connected layers, specifically fc6 and fc7.

The distinction matters for classification of the literature. A WSDDN-like detector is a two-stream MIL detector. A WSBDN, in the sense used by the 2019 end-to-end model, incorporates a basic detection head with both classification and box regression, and learns that head jointly with MIL. Collaborative frameworks such as WSCDN are therefore adjacent rather than identical: they inherit WSDDN-style weak proposal scoring, but reframe the problem around detector consistency between weak and strong branches. On VOC 2007, the collaborative weak detector $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$6 improves over the initial weak baseline $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$7 from $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$8 mAP and $$\mathcal{L}_{det} = \mathcal{L}_{cls} + \lambda\mathcal{L}_{loc},$$9 CorLoc to $\mathcal{L}_{cls}$0 mAP and $\mathcal{L}_{cls}$1 CorLoc, while the collaborative strong detector $\mathcal{L}_{cls}$2 reaches $\mathcal{L}_{cls}$3 mAP and $\mathcal{L}_{cls}$4 CorLoc; on VOC 2012, $\mathcal{L}_{cls}$5 reaches $\mathcal{L}_{cls}$6 mAP and $\mathcal{L}_{cls}$7 CorLoc, and $\mathcal{L}_{cls}$8 reaches $\mathcal{L}_{cls}$9 mAP and $\mathcal{L}_{loc}$0 CorLoc. These results contextualize the broader movement toward jointly optimized weakly supervised detectors, but they do not constitute a direct evaluation of WSBDN as a named method (Wang et al., 2018).