UpConvMixer Blocks in Satellite Detection

• UpConvMixer blocks are decoder units in GLOD that robustly upsample and refine features using techniques like PixelShuffle to preserve spatial fidelity.
• They integrate asymmetric fusion, depthwise atrous and pointwise convolutions, Highway gating, and CBAM attention for effective feature mixing.
• Empirical evaluations show improved PSNR and mAP metrics in high-resolution satellite imagery detection compared to traditional upsampling methods.

UpConvMixer blocks are the decoder and upsampling units in GLOD, a transformer-first architecture for object detection in high-resolution satellite imagery. In that system, four cascaded UpConvMixer blocks sit between a Swin Transformer encoder and a multi-path detection head, progressively recovering spatial resolution while refining features through U-Net-style skip connections, asymmetric fusion, separable ConvMixer-style mixing, CBAM attention, Highway gating, and PixelShuffle upsampling. The blocks are described as a form of “robust upsampling” because PixelShuffle increases resolution without checkerboard artifacts common to deconvolution, depthwise atrous convolution preserves local detail with enlarged receptive fields, Highway gating maintains information flow across a repeated mixer stack, and CBAM filters salient channels and spatial locations before the upsample step (Drapier et al., 15 Jul 2025).

1. Architectural position within GLOD

Within GLOD, UpConvMixer (UCM) blocks constitute the decoder stage. The Swin Transformer produces hierarchical features, and four cascaded UCMs decode from low-resolution, high-semantic maps back toward higher resolution. At each stage, a UCM concatenates a Swin-stage skip feature with the propagated decoder feature, with both tensors already aligned to the same spatial resolution by the surrounding neck (Drapier et al., 15 Jul 2025).

This placement makes the UCM the primary bridge between transformer-derived global semantics and the detection head. The downstream head is CenterNet-inspired and consumes fused features to generate heatmaps, offsets, and sizes. The paper states that the improved spatial fidelity from UCM is reflected in higher PSNR and mAP metrics. This suggests that the decoder is not merely reconstructive; it is directly implicated in preserving the localization sharpness needed for tiny-object detection in high-resolution satellite imagery (Drapier et al., 15 Jul 2025).

A common simplification is to regard UpConvMixer as an isolated upsampler. In GLOD, however, it operates jointly with Fusion Blocks. Outputs from adjacent UCMs are fused asymmetrically, from lower resolution to higher resolution, through pointwise projection and bilinear upsampling, after which the fused representation proceeds to the detection heads. Accordingly, UCM handles stagewise feature refinement and upsampling, while the Fusion Block strengthens cross-scale aggregation on top of UCM outputs (Drapier et al., 15 Jul 2025).

2. Internal composition of a single block

A single UCM typically receives two tensors, $$X_1 \in \mathbb{R}^{B \times C_1 \times H \times W}$$ and $$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$, already aligned to the same spatial resolution. These are concatenated along the channel dimension to form $$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$, where $C_{\mathrm{cat}} = C_1 + C_2$ (Drapier et al., 15 Jul 2025).

The concatenated tensor first passes through an Asymmetric Fusion front-end. This consists of three parallel convolutional paths with anisotropic kernels $1 \times 3$, $3 \times 3$, and $3 \times 1$, each followed by batch normalization and ReLU. Their outputs are summed and activated again:

$$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$

The stated purpose is to enhance directional structures, especially horizontal and vertical structure, before mixing (Drapier et al., 15 Jul 2025).

After asymmetric fusion, the block applies a mixer stack repeated $N=3$ times. Each repetition contains a depthwise atrous convolution for spatial mixing, followed by a pointwise $1 \times 1$ convolution for channel mixing. The depthwise stage operates on a per-channel kernel with dilation $$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$0, and the pointwise stage projects to a target channel count $$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$1. A Highway gate then combines the new representation with the previous state:

$$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$2

$$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$3

After the final repetition, CBAM is applied at the end of the block, and PixelShuffle with scale $$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$4 produces the upsampled output $$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$5 (Drapier et al., 15 Jul 2025).

The recommended hyperparameters are explicit: $$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$6 mixer repetitions, kernel size $$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$7, dilation $$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$8 for atrous depthwise convolution, BN after depthwise and pointwise convolutions, GELU activation, CBAM reduction ratio $$X_2 \in \mathbb{R}^{B \times C_2 \times H \times W}$$9, scale factor $$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$0, a Highway gate implemented via $$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$1 convolution plus sigmoid, and “same” padding for all spatial convolutions (Drapier et al., 15 Jul 2025).

3. Mathematical operators and tensor constraints

The UCM imposes a clear channel constraint before PixelShuffle. The pointwise stage expands or contracts to $$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$2, chosen to satisfy $$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$3. With $$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$4, this becomes $$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$5. This constraint ensures that PixelShuffle can rearrange channels into spatial dimensions without additional learned interpolation (Drapier et al., 15 Jul 2025).

The depthwise convolution is defined channelwise with dilation:

$$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$6

where $$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$7, $$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$8, and zero-padding is used at boundaries. The pointwise mixing is then

$$X = \mathrm{Cat}(X_1, X_2) \in \mathbb{R}^{B \times C_{\mathrm{cat}} \times H \times W}$$9

with $C_{\mathrm{cat}} = C_1 + C_2$0, followed by BN and GELU (Drapier et al., 15 Jul 2025).

CBAM is applied after the final mixer repetition. Channel attention is computed from average and max pooled summaries over spatial dimensions:

$C_{\mathrm{cat}} = C_1 + C_2$1

and spatial attention is computed from average and max pooled channel descriptors:

$C_{\mathrm{cat}} = C_1 + C_2$2

The output is

$C_{\mathrm{cat}} = C_1 + C_2$3

The PixelShuffle output is then defined elementwise by

$C_{\mathrm{cat}} = C_1 + C_2$4

for $C_{\mathrm{cat}} = C_1 + C_2$5 (Drapier et al., 15 Jul 2025).

The implementation notes in the paper also clarify a practical issue with Highway semantics: if strict Highway semantics require matching channels for the carried tensor $C_{\mathrm{cat}} = C_1 + C_2$6 across iterations, one may project $C_{\mathrm{cat}} = C_1 + C_2$7 once to $C_{\mathrm{cat}} = C_1 + C_2$8 via a $C_{\mathrm{cat}} = C_1 + C_2$9 convolution before the loop or initialize $1 \times 3$0 on the first iteration. The provided pseudocode shows the latter as a pragmatic lift to the required channel count (Drapier et al., 15 Jul 2025).

4. Efficiency profile and rationale for “robust upsampling”

The efficiency claims for UCM derive from both parameterization and operator choice. Ignoring BN affine terms and biases, the Asymmetric Fusion front-end has

$1 \times 3$1

parameters, obtained from the three anisotropic branches. One depthwise convolution contributes $1 \times 3$2 parameters, one pointwise convolution contributes $1 \times 3$3, and the Highway gate contributes $1 \times 3$4. The CBAM channel MLP contributes approximately

$1 \times 3$5

with $1 \times 3$6, and the spatial $1 \times 3$7 convolution contributes $1 \times 3$8 parameters. For one mixer iteration, excluding CBAM placed after the repeated stack,

$1 \times 3$9

For $3 \times 3$0 iterations, the total is

$3 \times 3$1

FLOPs scale linearly with $3 \times 3$2, with depthwise convolution contributing approximately $3 \times 3$3 and pointwise convolution approximately $3 \times 3$4 (Drapier et al., 15 Jul 2025).

The paper’s characterization of the block as robust is specific. PixelShuffle is said to avoid checkerboard artifacts common to transposed convolution. Depthwise atrous convolution preserves local detail while enlarging receptive fields without downsampling. Highway gating maintains information flow across the repeated mixer stack. CBAM filters salient channels and spatial locations before the upsample step. Together, these reduce aliasing and help recover tiny objects in high-resolution satellite imagery (Drapier et al., 15 Jul 2025).

The paper further states that DWConv plus PWConv is significantly cheaper than a full $3 \times 3$5 convolution at $3 \times 3$6 channels, which would be approximately

$3 \times 3$7

FLOPs. PixelShuffle has zero parameters and negligible FLOPs because it is a rearrangement operator. Highway combination and sigmoid are described as minor compared with the convolutional cost (Drapier et al., 15 Jul 2025).

5. Relation to other upsampling and pyramid designs

The paper contrasts UCM with several baseline upsamplers and feature-pyramid mechanisms. Against transposed convolution, the comparison emphasizes that deconvolution is learnable but prone to checkerboard patterns, heavier in parameters, and liable to blur small objects. UCM’s PixelShuffle plus attention is presented as avoiding these artifacts while preserving edges (Drapier et al., 15 Jul 2025).

Against bilinear upsampling plus convolution, the paper notes that interpolation can attenuate peaks in heatmaps and lacks learnable spatial rearrangement. UCM instead uses learned channel-to-space mapping before PixelShuffle, which the paper states preserves compact activations for tiny objects. Against standalone PixelShuffle, the limitation identified is that without spatial and channel mixing it can propagate noisy features; UCM adds anisotropic fusion, separable mixing, attention, and gating to produce cleaner salient detail (Drapier et al., 15 Jul 2025).

The comparison extends to ConvMixer-based upsamplers, FPN or PAFPN, and CARAFE. ConvMixer-style upsamplers provide depthwise and pointwise mixing but typically lack integrated fusion, attention, gating, and an explicit upsample mapping; UCM extends ConvMixer with a robust fusion front-end, CBAM, Highway, and PixelShuffle. FPN and PAFPN improve scale robustness but still rely on interpolation and standard convolutions at each level. CARAFE is described as content-aware and powerful but higher in memory and compute; UCM is presented as simpler and lighter while retaining robustness through attention and gating (Drapier et al., 15 Jul 2025).

The domain-specific rationale is also explicit. Anisotropic fusion kernels capture linear structures such as roads and runways; atrous depthwise convolution enlarges the receptive field without pooling; CBAM suppresses clutter; PixelShuffle avoids aliasing. The paper states that these properties help recover tiny vehicles and containers while preserving boundaries. This suggests that UCM is specialized for the texture statistics and geometric regularities of high-resolution satellite imagery rather than being introduced as a purely generic decoder primitive (Drapier et al., 15 Jul 2025).

6. Empirical behavior, training context, and practical limitations

On xView with fixed input size $3 \times 3$8, GLOD achieves $3 \times 3$9 mAP50 and $3 \times 1$0 mAP75, outperforming FPN-50 RFL at $3 \times 1$1 mAP50 and transformer baselines such as DQ-DETR at $3 \times 1$2 mAP50. For car detection heatmaps, GLOD reaches $3 \times 1$3 dB PSNR versus DNTR’s $3 \times 1$4 dB, which the paper interprets as sharper, less noisy peaks consistent with robust upsampling (Drapier et al., 15 Jul 2025).

The Fusion Block ablation isolates the effect of multi-scale aggregation on top of UCM outputs: with Fusion Blocks, performance is $3 \times 1$5 mAP50 and $3 \times 1$6 mAP75; without them, performance drops to $3 \times 1$7 mAP50 and $3 \times 1$8 mAP75. By contrast, the paper does not report a direct UpConvMixer replacement study against alternatives such as bilinear plus convolution or deconvolution. Accordingly, any claim that UCM alone accounts quantitatively for the full gain would exceed the reported evidence. The paper instead states that qualitative and PSNR gains suggest UCM’s upsampling and mixing improve small-object recovery and boundary fidelity (Drapier et al., 15 Jul 2025).

The training and inference settings that affect UCM are also specified. Optimization uses AdamW with learning rate $3 \times 1$9, $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$0, $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$1, $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$2, and cosine scheduling with warm restarts every $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$3 epochs, for $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$4k steps. The per-device batch size is $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$5, and gradient checkpointing enables large-resolution inputs of $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$6. Augmentation includes grayscale, solarization with threshold $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$7, histogram equalization, and random flips, while aggressive geometries, MixUp, and Mosaic are avoided to preserve the spatial priors exploited by the Transformer and UCM. Losses are CenterNet focal classification loss with $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$8 and $$AF_X = \sigma_{\mathrm{ReLU}}\!\big( BN_{1\times3}(Conv_{1\times3}(X)) + BN_{3\times3}(Conv_{3\times3}(X)) + BN_{3\times1}(Conv_{3\times1}(X)) \big).$$9, plus Smooth-L1 losses for offsets and sizes, all weighted equally; the minimum positive radius is $N=3$0, and the negative sampling ratio is $N=3$1 (Drapier et al., 15 Jul 2025).

The stated limitations and adaptation guidance delimit the scope of the block. Class imbalance and label noise, including clouded regions, still affect supervision, and UCM cannot compensate for missing labels. Although depthwise/pointwise convolutions and PixelShuffle are efficient, the overall transformer-first pipeline remains heavier than pure CNNs and requires careful batching and checkpointing. For adaptation to other resolutions, the paper recommends setting PixelShuffle scale $N=3$2 to the desired upsampling factor and ensuring $N=3$3 before shuffle. For domains beyond satellite imagery, it recommends retaining Asymmetric Fusion for directional structures, tuning the dilation schedule for domain-specific textures, and possibly increasing the CBAM reduction ratio to $N=3$4 for larger channel counts. If training stagnates with deeper mixer stacks, the authors recommend Highway gating over residual connections, reporting approximately $N=3$5 loss reduction during training (Drapier et al., 15 Jul 2025).