Binary neural networks (networks whose weights and activations are restricted to $\pm 1$, so that multiplication becomes a bitwise XNOR followed by counting the ones, called a popcount) are among the cheapest ways to run vision models on small devices. They fail worst on dense prediction tasks such as semantic segmentation and super-resolution. In these tasks the label of a pixel often depends on context far away from it, and a binary network built from $3 \times 3$ kernels sees only a small neighborhood unless it downsamples aggressively and discards boundaries. Every published attempt to give binary networks a global view keeps some part of the global operator out of 1-bit arithmetic. One work keeps a floating-point scan core, another decomposes attention into several binary bases, and a third multiplies attention by real-valued lookup scalars.
We propose a simpler kind of global operator that fits the binary regime cleanly. It is a depthwise one-dimensional convolution whose kernel is about twice the width or height of the feature map, with weights constrained to $\pm 1$ and zero padding at the borders. One horizontal layer followed by one vertical layer gives every pixel a receptive field covering the whole image. Each output is a single popcount over roughly $2W$ bits. Our central bet is that such a long kernel is easier to binarize than the local and attention operators the field has tried so far, for two reasons. Its large fan-in gives each output a fine integer resolution. And because it is static, it never needs the token magnitudes that binary attention destroys.
Core Claim
Large-kernel CNNs such as ParCNetV2 obtain a global receptive field from 1D depthwise kernels about twice the feature size (for example $111 \times 1$ on a $56 \times 56$ map). Their zero padding also embeds position implicitly. In full precision this reaches 84.6% ImageNet top-1 and 52.6 COCO box AP for the tiny variant. Oversized kernels have never been binarized. We claim that a binary version is a true end-to-end 1-bit global operator and that it closes a meaningful part of the dense-prediction gap. For channel $c$ at position $i$ along one axis, the binary oversized convolution is
$$y_c(i) = \alpha_c \sum_{k=-K/2}^{K/2} w_c(k)\, x_c(i+k),$$
where $w_c(k) \in \{-1,+1\}$ are the binarized kernel weights, $x_c \in \{-1,+1\}$ are the binarized activations (set to zero outside the map), $K \approx 2W$ is the kernel length, and $\alpha_c$ is a per-channel real scale folded into the next layer. The inner sum is one XNOR-popcount over the in-bounds taps.
Two properties follow directly. First, the output takes about $K$ distinct integer values, whereas a binary depthwise $3 \times 3$ kernel yields only 9. This is the representational bottleneck that XNOR-Net and Structured Binary Neural Networks identify. Second, the number of in-bounds taps depends on $i$, so the range of the popcount itself carries absolute position. This is the binary analogue of ParCNetV2's padding-induced position encoding, and it requires no coordinate channels.
The operator is deliberately content-independent. BVT-IMA shows that binary attention fails because every binary token has the same norm $\sqrt{d}$, which pushes attention toward background tokens. A static kernel sidesteps that failure mode entirely. It gives up input-dependent routing in exchange for staying inside pure bitwise logic.
Research Program
We will build a binary segmentation backbone in which the second half of each stage replaces $3 \times 3$ binary convolutions with paired horizontal and vertical binary oversized kernels. Following ParCNetV2, the kernel length scales with the stage resolution. Three pieces are reused from prior work. Bi-Real Net supplies real-valued identity shortcuts and its polynomial gradient approximation. BBCU supplies residual alignment, a learnable scale on the residual path in place of batch normalization, which that paper found harmful for dense restoration; we need it because the popcount range here is much wider than BBCU's $[-15, 15]$. BiT supplies elastic learnable thresholds and stepping-stone distillation from full precision through W1A2 to W1A1.
The analysis asks three questions. Does the binary kernel still encode position? We test this by probing absolute coordinates from features and by swapping zero padding for circular padding. Does the learned sign pattern implement an effective distance falloff? A binary kernel cannot taper in magnitude, only in sign density, so we measure the effective receptive field directly. Does large fan-in, rather than long reach, explain any gain? This last question is the key confound, and we isolate it with a matched-fan-in local kernel.
Deployment is part of the claim. Row-wise binary correlation is a bit-packed sliding popcount. That matters because ParCNetV2 notes that oversized kernels lack optimized library support, and MBU-Net notes that NVIDIA removed native bit matrix multiply-accumulate instructions on Hopper GPUs. We will therefore report measured CPU latency with bit-packed kernels alongside counts of binary operations.
Baselines
The naive reference is a Bi-Real/ReActNet-style binary ResNet segmentation network with only $3 \times 3$ kernels, trained with the same shortcuts, residual alignment and distillation as our model. It shows what binary context looks like with no global operator. Beating it by a clear mIoU margin on 19-class Cityscapes, at equal binary-operation budget, is the minimum bar.
The strongest existing fully binary context module is the binary dilated ASPP of Binary DAD-Net. It is also the only fully binary dense network in the file with a context module. It reports 96.60% mIoU on binary Cityscapes road detection and 95.25% on KITTI Road. We reimplement it on the same backbone, data, schedule and hyperparameter search budget, and we evaluate it on both the binary road task and the full 19-class task. For 19-class Cityscapes and ADE20K, no reported result for any binary network was found on Lacuna. Our comparisons there are therefore head-to-head reimplementations. The full-precision ceilings are those of CCNet: 81.9% test mIoU on Cityscapes and 45.76% on ADE20K. The file also lists the Dilated10 reference, 67.1% on Cityscapes, from the Cityscapes paper. Beating ASPP means a higher 19-class mIoU at equal or lower operations and model size.
For super-resolution, the strongest reported binary result is FRBC from Flexible Residual Binarization, at 25.61 dB PSNR on Urban100 ×4 against ReActNet's 25.35 dB. Urban100 is full of repeated building structure far apart in the image, which makes it a fair test of long-range context. Adding our operator to a BBCU-style binary SR network should beat 25.61 dB without FRB's second-order residual binarization, which doubles the binary work per layer.
Two ablations isolate the key ingredient. The first replaces the oversized kernel with a binary depthwise 2D kernel of the same fan-in but local support, about $11 \times 11$ for a 111-tap kernel. It separates "more taps" from "longer reach." The second keeps the oversized kernel but uses circular padding, which removes the implicit position signal. For classification context, we also report ImageNet top-1 against BHViT, whose small variant reports 70.1%, while noting that BHViT uses multi-basis attention.
Why not just…?
Why not just stack dilated binary convolutions? This is the Binary DAD-Net recipe. Its strong numbers come from a two-class road task, and dilated grids skip pixels. Binary features cannot fill those gaps with intermediate values, which worsens gridding. Dilation also covers the image only at fixed rates. Still, this is a real competitor, and our 19-class head-to-head at matched operations is exactly the experiment that settles it. If ASPP matches us, the oversized kernel is not worth its unusual implementation.
Why not just add a binary global-pooling broadcast? In the style of BCN or a squeeze-excitation branch, this costs almost nothing and gives every pixel global information. The objection is serious because pooling is so cheap. However, pooling produces one vector per image that carries no spatial layout. Segmentation needs "what is in the same row or column, and where," so we expect pooling to help large classes but not thin or boundary-dependent ones. We include it as a baseline and break mIoU down by class size to test this prediction.
Why not just keep the global operator in full precision? This follows Binarized Mamba-Transformer, which keeps its scan core in 32-bit floating point. It will probably be more accurate. We run a floating-point oversized-kernel variant inside the otherwise binary network as an upper reference. The question is what fraction of that accuracy the pure binary version keeps, and at what fraction of its latency on bit-packed hardware. If the binary kernel retains little of the gain, the honest conclusion is that global operators should stay multi-bit.
Why not just use binary criss-cross or other binary attention? Binary attention inherits the norm-collapse problem diagnosed by BVT-IMA. BVT-IMA's own fix still trails full precision by about 32 points on DeiT-Small. We include a binarized criss-cross attention baseline using BiT's two-set binarization, to show whether input-dependent routing survives 1-bit arithmetic at all in dense tasks.
Proposed Experiments
The main experiment trains all models on 19-class Cityscapes and on ADE20K at matched binary-operation budgets and reports mIoU, model size and measured bit-packed CPU latency. It also reports the binary Cityscapes road task and KITTI Road, which connect our models to published binary numbers. The secondary experiment is Urban100, Set5, Set14, B100 and Manga109 ×4 super-resolution against the FRBC and BBCU baselines.
The decisive checks are three. First, the binary oversized kernel must beat binary ASPP and binary global pooling on 19-class mIoU. Second, it must beat the matched-fan-in local kernel, which shows that reach and not only resolution matters. Third, circular padding must measurably hurt, and coordinate probes must fail without zero padding, which confirms the popcount-range position signal. A kernel-length sweep from $0.5W$ to $2W$ should show gains that saturate near full coverage. If instead the gains track the tap count regardless of reach, our explanation is wrong and the contribution reduces to "wider binary kernels."
References
- ParCNetV2: Oversized Kernel with Enhanced Attention
- XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks
- Structured Binary Neural Networks for Accurate Image Classification and Semantic Segmentation
- BVT-IMA: Binary Vision Transformer with Information-Modified Attention
- Bi-Real Net: Enhancing the Performance of 1-Bit CNNs with Improved Representational Capability and Advanced Training Algorithm
- Basic Binary Convolution Unit for Binarized Image Restoration Network
- BiT: Robustly Binarized Multi-distilled Transformer
- Zeros can be Informative: Masked Binary U-Net for Image Segmentation on Tensor Cores
- Binary DAD-Net: Binarized Driveable Area Detection Network for Autonomous Driving
- CCNet: Criss-Cross Attention for Semantic Segmentation
- The Cityscapes Dataset for Semantic Urban Scene Understanding
- Flexible Residual Binarization for Image Super-Resolution
- BHViT: Binarized Hybrid Vision Transformer
- Broadcasting Convolutional Network for Visual Relational Reasoning
- Binarized Mamba-Transformer for Lightweight Quad Bayer HybridEVS Demosaicing