--- kind: concept slug: dynamic-convolution title: Dynamic convolution summary: >- A convolution whose filters are generated per location from the local input, instead of one learned filter set shared everywhere. The one compiled variant (DHiF) adds a tanh-bounded generated filter bank to the input before a standard convolution. Author-reported IoU gains in infrared small target detection, at a small speed cost. tags: [convolution, dynamic-filters, infrared-small-target-detection] papers: ["2602.02969"] related: [convolutional-neural-network, deformable-convolution, hypernetwork, frequency-domain-learning] status: draft updated_at: 2026-11-06 --- ## Overview An ordinary convolution slides a window over the input and applies one learned filter set at every position. Dynamic convolution keeps the sliding window but generates the filters from the input at each position, so the filters adapt to local content [arxiv:2602.02969 §II-B.1]. The generated filters are produced by a small learned layer, so the block can be built from standard operations and placed where a convolution would go [arxiv:2602.02969 §II-B.3]. Only one compiled card covers this concept so far, so this page describes one variant. Other dynamic-convolution work is not yet compiled here and is not covered by the evidence below. ## Variants ### DHiF: per-location high-frequency filter bank The paper calls the block DHiF (dynamic high-frequency convolution). For each window it does the following [arxiv:2602.02969 §II-B.1] [arxiv:2602.02969 §II-B.3]: 1. normalises the local input (the paper names a normalisation layer but does not define it; see Notes for implementers); 2. projects the k² local values to k² filter vectors of length k² with a linear map and bounds every entry with tanh [arxiv:2602.02969 eq. 5] [arxiv:2602.02969 eq. 6]; 3. applies that k² × k² bank to the local patch of every channel [arxiv:2602.02969 eq. 7]; 4. adds the filtered response to the original patch and forms the output with per-filter weights [arxiv:2602.02969 eq. 8]. The motivation is that a filter should respond to high-frequency detail rather than flat regions. The paper argues that kernels with positive and negative entries and a near-zero sum do this, and it enforces this by bounding each entry to (-1, 1) [arxiv:2602.02969 §II-B.2]. The bound is the only mechanism in the implementation; the near-zero sum is not enforced. In the paper's residual block, DHiF replaces the first convolution, followed by batch norm and an activation, and the standard convolution comes after it [arxiv:2602.02969 supp. §II-B] [arxiv:2602.02969 supp. eq. 1]. The variant is evaluated on single-frame infrared small target detection with seven convolution-based networks: CNN and CNN-Transformer hybrids [arxiv:2602.02969 §III-B]. Card: [Dynamic High-frequency Convolution for Infrared Small Target Detection](/papers/2602.02969). #### Reported results (author-reported) Each row of the main table is one network with its encoder convolutions replaced; each cell is a single reported value, with no variance given [arxiv:2602.02969 tab. 1]. For DNANet: | Dataset | Conv | IoU | nIoU | P_d | F_a | |---|---|---|---|---|---| | IRSTD-1K | Standard | 63.27 | 63.56 | 88.55 | 1.25 | | IRSTD-1K | DHiF | 69.42 | 65.68 | 90.57 | 1.38 | | NUAA-SIRST | Standard | 71.97 | 78.66 | 95.04 | 6.31 | | NUAA-SIRST | DHiF | 77.68 | 79.03 | 95.42 | 1.53 | Source: [arxiv:2602.02969 tab. 1]. Units: IoU, nIoU and P_d in 10^-2; F_a in 10^-5 [arxiv:2602.02969 §III-A]. Across the 14 network-and-dataset rows, DHiF has higher IoU than the standard convolution in every row. The IoU gain ranges from +0.09 (MSHNet, NUAA-SIRST) to +40.24 (FC3Net, NUAA-SIRST), where the standard row is an outlier the paper does not explain [arxiv:2602.02969 tab. 1]. Ablations, all on NUAA-SIRST with DNANet unless noted: - **Kernel size.** DHiF has higher IoU than the standard convolution at k = 3 (77.68 vs 71.97), k = 5 (76.87 vs 72.36) and k = 7 (76.80 vs 74.17); the best DHiF IoU is at k = 3 [arxiv:2602.02969 supp. tab. II]. - **Bound on the filters.** With no non-linearity, GELU, LeakyReLU or sigmoid, IoU is 71.81, 71.12, 72.92 and 72.25. With tanh it is 77.68 [arxiv:2602.02969 supp. tab. III]. GELU is below the no-activation baseline; LeakyReLU and sigmoid are above it; all three are below tanh. - **Position in the block.** DHiF before the standard convolution gives 77.68 IoU; the reverse order gives 74.55 [arxiv:2602.02969 supp. tab. I]. - **Noise.** With salt-and-pepper noise on DNANet, IoU is 66.48 for the standard convolution and 72.08 for DHiF [arxiv:2602.02969 supp. tab. V]. - **Level.** Replacing only encoder level 2 (two blocks) gives IoU 78.37; only level 5 gives 75.48 [arxiv:2602.02969 tab. 2]. The main experiments use levels 2 and 3 [arxiv:2602.02969 §III-C.2]; see Contradictions. Cost on DNANet at 512 × 512: FPS 8.61 → 8.02 (about 6.9% fewer frames per second), parameters 17.92 M → 17.93 M, and runtime 0.1161 s → 0.1247 s [arxiv:2602.02969 supp. tab. IV]. The paper reports no peak-memory figure. ## When to use which The evidence is one paper on one task, with single-run numbers. Use the following as a guide, not a guarantee. | Situation | Use DHiF? | Why | |---|---|---| | Small-target detection in single infrared frames, convolutional encoder | Candidate; test on your data | IoU rose in all 14 rows of [arxiv:2602.02969 tab. 1]; the paper's data is the only evidence | | Inference speed is a hard budget | Measure first | DNANet FPS fell from 8.61 to 8.02 at 512 × 512 [arxiv:2602.02969 supp. tab. IV]; the paper did not measure other hardware or memory | | Kernel size larger than 3 | Expect a smaller gain | At k = 5 and k = 7 the gain on NUAA-SIRST is smaller than at k = 3 [arxiv:2602.02969 supp. tab. II]; k = 3 is the paper's default | | Replacing the input (stem) layer | Avoid unless tested | The paper reports no gain and lower P_d for DHiF at the input layer [arxiv:2602.02969 §III-C.1] | | Non-convolutional backbone | Not covered | The method is designed for convolutional structures [arxiv:2602.02969 §IV] | | Domains other than infrared small targets | Not covered | Only the two SIRST datasets were tested [arxiv:2602.02969 §III-A] | ## Implementation sketch The sketch follows the paper's equations [arxiv:2602.02969 eq. 5]-[arxiv:2602.02969 eq. 8]. Channel normalisation is left as a choice, as the paper leaves it open. ```pseudocode # DHiF forward pass for one feature map [arxiv:2602.02969 §II-B.3] # Input: x (B, C, H, W); kernel size k # Output: y (B, C_out, H, W) def dhif(x, k, generator, out_weight): # Gather the k x k window at every location [§II-B.1] patches = unfold(x, k) # (B, C, H, W, k*k) # Reduce the channels to one value per tap; the paper leaves this open local = normalise_over_channels(patches) # (B, H, W, k*k) # Generate one k*k x k*k filter bank per location [eq. 5] # Bound every entry to (-1, 1) with tanh [eq. 6] bank = tanh(generator(local)) # (B, H, W, k*k, k*k) # Apply each location's bank to every channel's patch [eq. 7] filtered = apply_bank(patches, bank) # (B, C, H, W, k*k) # Add the input patch to the filtered response [eq. 8] combined = patches + filtered # (B, C, H, W, k*k) # Weighted sum over channels and taps, once per output filter [eq. 8] return weighted_sum(combined, out_weight) # (B, C_out, H, W) ``` In the residual block, the DHiF output is followed by batch norm and an activation, then the standard convolution [arxiv:2602.02969 supp. eq. 1]. The shortcut path is not specified in the paper's equation. ### Notes for implementers - **Normalisation is unspecified.** The paper names a normalisation layer but does not define it [arxiv:2602.02969 §II-B.3]. The card for this paper records that the public code uses an L2 norm over channels; see its Paper vs code section. Results may depend on this choice. - **The sum over taps and channels.** Equation 8 is written as a sum over k² only, but the weights have shape C × k² [arxiv:2602.02969 eq. 8]. Sum over channels and taps, or the output shape will be wrong. - **Cost scales with k⁴ per pixel.** Each location builds a k² × k² bank, and the unfold step materialises a copy of the feature map [arxiv:2602.02969 supp. tab. IV]. Memory is not measured in the paper. - **Attention in the authors' block.** The card records that the authors' DNANet block applies channel and spatial attention before the residual addition, which the paper's block equation does not show [arxiv:2602.02969 supp. eq. 1]. The paper's numbers may depend on this. ## Contradictions No disagreement between papers: one compiled card covers this concept. Within the paper and between the paper and its public code, these points differ: - **Levels replaced in DNANet.** Table 2 lists levels 2 to 5, the text names levels 2 and 3, and the repository uses levels 2 and 3 [arxiv:2602.02969 tab. 2] [arxiv:2602.02969 §III-C.1] [arxiv:2602.02969 §III-C.2]. The main-table IoU of 77.68 matches no single row of the level table. - **Learning-rate schedule.** The paper says the rate halves after every 50 epochs; the code's first halving is at epoch 100 [arxiv:2602.02969 supp. §III-B]. - **Initialisation.** The paper cites Kaiming initialisation; the code uses the PyTorch default for convolutions [arxiv:2602.02969 supp. §III-B]. - **Input layer.** The paper advises against DHiF at the input layer, while its public code uses it there for one network [arxiv:2602.02969 §III-C.1]. ## Papers - [Dynamic High-frequency Convolution for Infrared Small Target Detection](/papers/2602.02969): introduces DHiF, a convolution with per-location tanh-bounded filter banks, and reports IoU gains in seven SIRST networks.