Deformable Convolution
A convolutional neural network operation that learns two-dimensional offsets at the standard convolution sampling points, allowing the kernel to adaptively focus on irregular geometric structures; bilinear interpolation is required to realize sampling at non-integer coordinates.
Timeline (last 90 days)
The bilinear interpolation and dynamic offset prediction in deformable convolution are relatively sensitive to quantization, and low-bit quantization may amplify interpolation error.
Qualcomm's Hexagon NPU (and the earlier DSP) has relatively limited operator support, mainly optimizing static-graph operators such as regular convolution, pooling and activation; for dynamic operators such as deformable convolution and grid_sample, execution typically needs to fall back to the CPU or GPU, or developers must manually implement approximate versions.
Mobile GPUs (such as the Adreno series) support deformable convolution somewhat better, because GPU shader programs can flexibly implement interpolation and dynamic indexing.
The BiRefNet architecture makes extensive use of deformable convolutions.
In deformable convolution, the offsets are predicted from the input feature map by an additional convolutional layer, and the sampling positions become non-integer coordinates, requiring bilinear interpolation to obtain values at the corresponding positions from the feature map.
Deformable convolution introduces data-dependent dynamic indexing and interpolation operations, breaking the static computation graph property of standard convolution; for mobile inference frameworks that rely on fixed operator mapping and static scheduling compilation, such dynamic operators often cannot be effectively graphed or hardware-accelerated.
Deformable convolution was first proposed by Dai Jifeng et al. from Microsoft Research Asia at the ICCV 2017 conference (DCNv1), with an improved version DCNv2 released in 2019.
Deformable convolution was proposed by Microsoft Research Asia in 2017 and adaptively focuses on irregularly shaped target regions by learning offsets for each sampling position of the convolution kernel.
All Facts (8)
The bilinear interpolation and dynamic offset prediction in deformable convolution are relatively sensitive to quantization, and low-bit quantization may amplify interpolation error.
50%UnverifiedQualcomm's Hexagon NPU (and the earlier DSP) has relatively limited operator support, mainly optimizing static-graph operators such as regular convolution, pooling and activation; for dynamic operators such as deformable convolution and grid_sample, execution typically needs to fall back to the CPU or GPU, or developers must manually implement approximate versions.
50%UnverifiedMobile GPUs (such as the Adreno series) support deformable convolution somewhat better, because GPU shader programs can flexibly implement interpolation and dynamic indexing.
50%UnverifiedThe BiRefNet architecture makes extensive use of deformable convolutions.
50%UnverifiedIn deformable convolution, the offsets are predicted from the input feature map by an additional convolutional layer, and the sampling positions become non-integer coordinates, requiring bilinear interpolation to obtain values at the corresponding positions from the feature map.
50%UnverifiedDeformable convolution introduces data-dependent dynamic indexing and interpolation operations, breaking the static computation graph property of standard convolution; for mobile inference frameworks that rely on fixed operator mapping and static scheduling compilation, such dynamic operators often cannot be effectively graphed or hardware-accelerated.
50%UnverifiedDeformable convolution was first proposed by Dai Jifeng et al. from Microsoft Research Asia at the ICCV 2017 conference (DCNv1), with an improved version DCNv2 released in 2019.
50%UnverifiedDeformable convolution was proposed by Microsoft Research Asia in 2017 and adaptively focuses on irregularly shaped target regions by learning offsets for each sampling position of the convolution kernel.
50%