Cross-Layer Feature Pyramid Transformer for Small Object Detection in Aerial Images

Modern object detectors typically decrease the resolution of the input image through successive layers of convolution and pooling, striving to achieve an optimal balance between performance and computational complexity [14, 15, 16]. Therefore, detecting small objects is inherently more challenging than common object detection, as their diminutive size increases the risk of information loss during downsampling.

For small object detection in aerial images, ClusDet [17] employs a coarse-to-fine scheme that initially detects dense object clusters and then refines the search within these clusters to enhance the model’s ability to detect small objects. DMNet [18] simplifies the training process of ClusDet by employing a density map generation network to produce density maps for cluster prediction. Following the similar detection pipeline, CRENet [19] and GLSAN [4] further enhance the clustering prediction algorithm and optimize the fine-grained prediction scheme. UFPMP-Det [6] employs the UFP module and MP-Net to predict sub-regions, assembling them into a single image for efficient single-inference, thereby achieving improved detection accuracy and efficiency. CEASC [20] utilizes sparse convolution to optimize conventional detectors for object detection in aerial images, reducing computational requirements while maintaining competitive performance. DTSSNet [21] introduces a manually designed block between the Backbone and Neck to enhance the model’s sensitivity to multi-scale features and incorporates a training sample selection method specifically for small objects.

The above solutions optimize various detectors for object detection scenarios in aerial images, whereas we propose a new feature pyramid network specifically tailored for small object detection in this context.

To alleviate the substantial computational costs of image pyramids, feature pyramid networks (FPNs) have emerged as an effective and efficient alternative that enhances the performance of various detectors. FPN [7] utilizes a series of top-down shortcut branches to augment the semantic information lacking in shallow feature maps. Based on FPN, PAFPN [12] proposes using bottom-up shortcut branches to address the deficiency of detailed information in deep feature maps. Libra-RCNN [22] refines original features by incorporating non-local blocks to obtain balanced interactive features. To mitigate the semantic gap in multi-scale feature maps, AugFPN [23] introduces the consistent supervision branch and proposes ASF for dynamic feature integration across multiple scales. FPG [8] represents the feature scale space using a regular grid fused with multi-directional lateral connections between parallel paths, thereby enhancing the model’s feature representation. AFPN [11] iteratively refines multi-scale features through cross-level fusion of deep and shallow feature maps, achieving competitive performance in object detection with common scale distribution.

Distinct from previous approaches, we propose CFPT, which leverages global contextual information and strategically emphasizes shallow feature maps to enhance the detection of small objects in aerial images.

As an extension of Transformer [24] in computer vision, Vision Transformer (ViT) [25] have demonstrated significant potential across various visual scenarios [26, 27, 28]. Due to the quadratic computational complexity of conventional ViTs with respect to image resolution, subsequent research has predominantly focused on developing lightweight alternatives. Swin Transformer [29] restricts interactions to specific windows and achieves a global receptive field by shifting these windows during the interaction process. Local ViTs [30, 31, 32] incorporate local inductive bias through interactions within local windows, effectively reducing the model’s computational complexity and accelerating its convergence speed. Axial Attention [33] reduces computational complexity by confining interactions to strips along the width and height of the image.

Following the similar lightweight concept, we design two attention blocks with linear complexity (i.e., CCA and CSA) to capture the global contextual information across layers along various directions (i.e., spatial-wise and channel-wise), thereby enhancing the model’s ability to detect small objects.

As illustrated in Fig. 4, CFPT employs multiple parallel CBR blocks to construct the input for cross-layer feature interaction using the multi-level feature map outputs from the feature extraction network (e.g., ResNet[34]), thereby reducing computational complexity and meeting the architectural requirements of most detectors. By leveraging stacked Cross-layer Attention Modules (CAMs), CFPT is able to enhance the model’s ability to utilize both global contextual information and cross-layer multi-scale information.

Specifically, the CAM module consists of a sequence of Cross-layer Channel-wise Attention (CCA) and Cross-layer Spatial-wise Attention (CSA). The CCA facilitates local cross-layer interactions along the channel dimension, consequently establishing a global receptive field along the spatial dimension through interactions within each channel-wise token group. Conversely, CSA facilitates local cross-layer interactions along the spatial dimension, capturing global contextual information along the channel dimension through interactions within each spatial-wise token group. In addition, we further improve the gradient gain by using shortcut branches between input and output of CAM.

Assume that the feature map of each scale after the CBR block can be represented as $X=\{X_{i}\in\mathcal{R}^{H_{i}\times W_{i}\times C}\}^{L-1}_{i=0}$, where $L$ is the number of input layers, and the spatial resolution $H_{i}\times W_{i}$ of each feature map increases with $i$ while maintaining the same number of channels $C$. The above process can be described as

where $Y=\{Y_{i}\in\mathcal{R}^{H_{i}\times W_{i}\times C}\}^{L-1}_{i=0}$ is a set of multi-scale feature maps after cross-layer interaction, maintaining the same shape as the corresponding input feature maps.

It is noteworthy that our CFPT eliminates the complex feature upsampling operations and layer-by-layer information transmission mechanisms, which are prone to information loss during inter-layer transmission and contribute to increased computational load and memory access delays. Instead, we perform a local grouping operation on multi-scale feature maps by utilizing the mutual receptive field size between scales, and subsequently facilitate information mixing between scales through a one-step cross-layer neighboring interaction operation. This approach enables features at each scale to acquire information from other layers in a balanced manner (even if the layers are distant) while facilitating self-correction and benefiting from the inductive bias provided by local interactions [32].

Suppose the set of input feature maps of the CCA is $X=\{X_{i}\in\mathcal{R}^{H_{i}\times W_{i}\times C}\}^{L-1}_{i=0}$. As shown in Fig. 5(a), CCA executes multi-scale neighboring interactions across layers along the channel dimension, thereby furnishing global contextual information along the spatial dimension for each channel-wise token. To construct the interactive input, we first perform Channel Reconstruction (CR) on the feature map of each scale to ensure they have the same spatial resolution, thereby obtaining $\hat{X}=\{\hat{X}_{i}\in\mathcal{R}^{H_{0}\times W_{0}\times\frac{H_{i}W_{i}}{H_{0}W_{0}}\times C}\}^{L-1}_{i=0}$. CR is an operator similar to Focus in YOLOv5, but it differs in that it does not use additional operations for feature mapping. Instead, CR stacks feature values from the spatial dimension into the channel dimension, thereby achieving consistent spatial resolution while maintaining efficiency. The above process can be described as

Next, we perform the Overlapped Channel-wise Patch Partition (OCP) to form channel-wise token groups, which can be viewed as the Patch Embedding [25] with overlapping regions in local areas along the channel dimension, where the patch sizes vary for feature maps at different scales. Specifically, according to the shape of multi-scale features, the channel sizes of adjacent feature maps in $\hat{X}$ differ by a factor of $4$ (i.e., $\frac{H_{i}W_{i}}{H_{0}W_{0}}=4^{i}$). To construct overlapping neighboring interaction groups, we introduce an expansion factor $\alpha$ to perform OCP on $\hat{X}$, resulting in $\tilde{X}=\{\tilde{X}_{i}\in\mathcal{R}^{C\times(4^{i}+\alpha)\times H_{0}W_{0}}\}^{L-1}_{i=0}$. The above process can be described as

Taking the feature map of the $i$-th layer as an example, after obtaining $\tilde{X}_{i}$, we employ the cross-layer consistent multi-head attention to capture the global dependencies along the spatial dimension, thereby obtaining the interaction result $\tilde{Y}_{i}\in\mathcal{R}^{C\times(4^{i}+\alpha)\times H_{0}W_{0}}$.

where $W^{q},W^{k}_{i},W^{v}_{i}\in\mathcal{R}^{C\times C}$ are the linear projection matrices. $K=[K_{0},K_{1},\ldots,K_{L-1}]\in\mathcal{R}^{C\times\sum_{i=0}^{L-1}(4^{i}+\alpha)\times H_{0}W_{0}}$ and $V=[V_{0},V_{1},\ldots,V_{L-1}]\in\mathcal{R}^{C\times\sum_{i=0}^{L-1}(4^{i}+\alpha)\times H_{0}W_{0}}$ represent the concatenated keys and values, respectively, where $[\cdot]$ represents concatenation operation. $P_{i}$ denotes the $i$-th Cross-layer Consistent Relative Positional Encoding (CCPE), with details will be introduced in Section III-D. Note that for simplicity, we only consider the case where the number of heads is $1$. In practice, we employ the multi-head mechanism to capture global dependencies for each channel-wise token.

After obtaining the interaction results $\tilde{Y}=\{\tilde{Y}_{i}\}^{L-1}_{i=0}$ for feature maps at each scale, we apply the Reverse Overlapped Channel-wise Patch Partition (ROCP) to restore the impact of OCP and obtain $\hat{Y}=\{\hat{Y}_{i}\in\mathcal{R}^{H_{0}\times W_{0}\times\frac{H_{i}W_{i}}{H_{0}W_{0}}\times C}\}^{L-1}_{i=0}$. As the reverse operation of OCP, ROCP aims to restore the original spatial resolution using the same kernel size and stride as OCP.

Finally, we use Spatial Reconstruction (SR) to obtain the result $Y=\{Y_{i}\in\mathcal{R}^{H_{i}\times W_{i}\times C}\}^{L-1}_{i=0}$ that matches the shape of the input $X$.

Similarly, let the set of input feature maps for the CSA be denoted as $X=\{X_{i}\in\mathcal{R}^{H_{i}\times W_{i}\times C}\}_{i=0}^{L-1}$. As illustrated in Fig. 5(b), CSA performs multi-scale neighboring interactions across layers along the spatial dimension, providing global contextual information along the channel dimension for each spatial-wise token.

Since the channel sizes of the input feature maps are matched after the CBR block(e.g., 256), there is no need to align their sizes using methods such as CR and SR, as is done in CCA. Therefore, we can directly perform Overlapped Spatial-wise Patch Partition (OSP) to form spatial-wise token groups, which can be viewed as sliding crops on feature maps of different scales using rectangular boxes of varying sizes. Suppose the expansion factor for OSP is $\beta$, through the above operation, we can get $\hat{X}=\{\hat{X}_{i}\in\mathcal{R}^{H_{0}\times W_{0}\times(2^{i}+\beta)^{2}\times C}\}_{i=0}^{L-1}$. The above process can be expressed as

Then, we can perform local interaction within cross-layer spatial-wise token groups and use the cross-layer consistent multi-head attention to capture global dependencies along the channel dimension, thereby obtaining $\hat{Y}=\{\hat{Y}_{i}\in\mathcal{R}^{H_{0}\times W_{0}\times(2^{i}+\beta)^{2}\times C}\}_{i=0}^{L-1}$. For the feature map of the $i$-th layer, this process can be expressed as follows:

where $W^{q},W^{k}_{i},W^{v}_{i}\in\mathcal{R}^{C\times C}$ are the linear projection matrices. $K=[K_{0},K_{1},...,K_{L-1}]\in\mathcal{R}^{H_{0}\times W_{0}\times\sum_{i=0}^{L-1}(2^{i}+\beta)^{2}\times C}$ and $V=[V_{0},V_{1},...,V_{L-1}]\in\mathcal{R}^{H_{0}\times W_{0}\times\sum_{i=0}^{L-1}(2^{i}+\beta)^{2}\times C}$. $P_{i}$ denotes the Cross-layer Consistent Relative Positional Encoding (CCPE) for the $i$-th layer.

Next, we employ Reverse Overlapped Spatial Patch Partition (ROSP) to reverse the effect of OSP and obtain the interaction result set $Y=\{Y_{i}\in\mathcal{R}^{H_{i}\times W_{i}\times C}\}_{i=0}^{L-1}$.

Since each token within their cross-layer token groups maintains specific positional relationships during the interaction process. However, the vanilla multi-head attention treats all interaction tokens uniformly, resulting in suboptimal results for position-sensitive tasks such as object detection. Therefore, we introduce Cross-layer Consistent Relative Positional Encoding (CCPE) to enhance the cross-layer location awareness of CFPT during the interaction process.

The main solution of CCPE is based on aligning mutual receptive fields across multiple scales, which is determined by the properties of convolution. Taking CSA as an example, the set of attention maps between each spatial-wise token group can be expressed as $A=\{A_{ij}\in\mathcal{R}^{N\times H_{0}\times W_{0}\times(2^{i}+\beta)^{2}\times(2^{j}+\beta)^{2}}\}_{i,j=0}^{L-1}$, where $N$ is the number of heads and $A_{ij}=Q_{i}K_{j}/\sqrt{d_{k}}$, as defined in Equation 9. For simplicity, we ignore $H_{0}$ and $W_{0}$ and define $s_{i}^{h}=s_{i}^{w}=2^{i}+\beta$ and $s_{j}^{h}=s_{j}^{w}=2^{j}+\beta$, where $(s_{i}^{h},s_{i}^{w})$ and $(s_{j}^{h},s_{j}^{w})$ represent the height and width of the spatial-wise token groups in the $i$-th and $j$-th layers. Therefore, the set of attention maps can be re-expressed as $A=\{A_{ij}\in\mathcal{R}^{N\times s_{i}^{h}s_{i}^{w}\times s_{j}^{h}s_{j}^{w}}\}_{i,j=0}^{L-1}$.

The process of CCPE is shown Fig. 6. We define a learnable codebook $C\in\mathcal{R}^{N\times(2s_{L-1}^{h}-1)\times(2s_{L-1}^{w}-1)}$ and obtain the relative position information between any two tokens from the codebook by calculating their cross-layer consistent relative position index. For simplicity, consider the interaction of the spatial-wise token groups from the $i$-th and $j$-th layers, where $P_{i}\in\mathcal{R}^{s^{h}_{i}\times s^{w}_{i}\times 2}$ and $P_{j}\in\mathcal{R}^{s^{h}_{j}\times s^{w}_{j}\times 2}$ represent their respective absolute coordinate matrices.

To obtain the relative position information of $P_{j}$ relative to $P_{i}$, we first centralize their coordinates using their respective spatial-wise token group sizes to obtain $\hat{P}_{i}$ and $\hat{P}_{j}$.

Then we can derive $\tilde{P}_{i}$ and $\tilde{P}_{j}$ by projecting their coordinates to the largest spatial-wise token group.

Subsequently, we can calculate the relative distance between $\tilde{P}_{i}$ and $\tilde{P}_{j}$ and convert the relative distance into the index of the codebook.

Finally, we extract the corresponding positional embedding vector from the codebook $C$ using the index $I_{ij}$ and superimpose it on the original attention map $A_{ij}$, resulting in the output $\hat{A}_{ij}$ with enhanced positional information.

where $f(a,b)$ is the bilinear interpolation function used to ensure the differentiability when extracting relative positional information from the codebook using non-integer coordinates derived from Equation 13, with $a$ representing the codebook and $b$ denoting the non-integer coordinates matrix.

In this section, we will analyze the computational complexity of both CCA and CSA. In addition, since the sizes of the spatial-wise and channel-wise token groups remain constant during both training and testing phases, their computational complexity scales linearly with the spatial resolution of the input feature maps.

We evaluate the effectiveness of the proposed CFPT by applying it to two challenging datasets specifically designed for small object detection from the drone’s perspective: VisDrone2019-DET [2] and TinyPerson [3].

We implement the proposed CFPT using PyTorch [48] and the MMdetection Toolbox [49]. All models are trained and tested on a single RTX 3090, with a batch size of 2. We use SGD as the optimizer for model training with a learning rate of 0.0025, momentum of 0.9, and weight decay of 0.0001. We conduct ablation studies and compare the performance of various state-of-the-art feature pyramid networks on the VisDrone2019-DET dataset, using an input resolution of $1333\times 800$ and 1$\times$ schedule (12 epochs). To accelerate the convergence of the model, the linear warmup strategy is employed at the beginning of training. For performance comparisons among various state-of-the-art detectors on the VisDrone2019-DET dataset, we train the models for 15 epochs to ensure full convergence following CEASC [20].

In our experiments on the TinyPerson dataset [3], we mitigate excessive memory usage by dividing the high-resolution images into evenly sized blocks with a $30\%$ overlap ratio. Each block is proportionally scaled to ensure that the shortest side measures 512 pixels. To comprehensively evaluate the model performance, we set the batch size to 1 with a 1$\times$ schedule for model training and employ both multi-scale training and multi-scale testing.

We initially compare the performance of the proposed CFPT with various state-of-the-art feature pyramid networks based on RetinaNet [1] on the VisDrone2019-DET dataset. As illustrated in Table I, our CFPT achieves the best results in RetinaNet across different backbone networks, including ResNet-18, ResNet-50, and ResNet-101, striking the optimal balance between performance and computational complexity. In addition, compared to SSFPN, which focuses on small object detection in aerial images, our CFPT achieves better performance (+0.8 AP, +0.5 AP, and +0.4 AP) with fewer parameters (-3.8 M, -3.5 M, and -3.5 M) and lower FLOPs (-55.5 G). This demonstrates the applicability of CFPT for small object detection in aerial images.

To further verify the effectiveness of CFPT, we replace the feature pyramid network in state-of-the-art detectors with CFPT and compare their performance on the VisDrone2019-DET and TinyPerson datasets.

We conduct the qualitative analysis of CFPT by visualizing the detection results on the VisDrone2019-DET and TinyPerson datasets, with the confidence threshold for all visualizations set to 0.3. As shown in Fig. 8, we apply CFPT to GFL and compare it qualitatively with the baseline model (i.e., GFL) and CEASC on VisDrone2019-DET dataset. The application of CFPT effectively reduces the model’s missed detection rate (first and third rows) and false detection rate (second row), thereby improving its overall performance. In addition, the third row of Fig. 8 demonstrates CFPT’s effectiveness in detecting small objects. As shown in Fig. 10, the detection results on the TinyPerson dataset further validate the above explanations, demonstrating that CFPT effectively reduces the missed detection and false detection rates while enhancing the model’s ability to detect small objects.