We present a simple, highly modularized network architecture for image classification. Our network is constructed by repeating a building block that aggregates a set of transformations with the same topology. Our simple design results in a homogeneous, multi-branch architecture that has only a few hyper-parameters to set. This strategy exposes a new dimension, which we call "cardinality" — the size of the set of transformations, as an essential factor in addition to the dimensions of depth and width. On the ImageNet-1K dataset, we empirically show that increasing cardinality is able to improve classification accuracy even under the setting of maintaining complexity. Moreover, increasing cardinality is more effective than going deeper or wider when we increase the capacity. Our model, named ResNeXt, is the foundation of our entry to the ILSVRC 2016 classification task in which we secured 2nd place.

Research on visual recognition is undergoing a transition from "feature engineering" to "network engineering". The design of network architectures has become increasingly important, as demonstrated by the progression from AlexNet to VGGNet to Inception and ResNet. While deeper and wider networks tend to improve accuracy, the design space is enormously large and hand-crafting architectures requires substantial expertise. Inception models demonstrate that carefully designed topologies can achieve compelling accuracy with lower theoretical complexity than VGGNet, but each Inception module involves many individually customized hyperparameters, making it difficult to adapt to new tasks.

The authors adopt VGGNet/ResNet's strategy of stacking building blocks of the same shape, which is simple and allows extension to any number of transformations. In contrast to Inception's heterogeneous branches, all paths in ResNeXt share identical topology. This leads to a key insight: the "split-transform-merge" strategy can be abstracted as aggregating transformations, where the number of transformations — the cardinality — becomes a concrete, measurable dimension orthogonal to depth and width. Experiments show that a 101-layer ResNeXt achieves better accuracy than a 200-layer ResNet with only 50% of its complexity.

The paper contextualizes ResNeXt among several lines of work. Multi-branch networks such as Inception use carefully customized branches with different filter sizes and pooling operations. Grouped convolutions, originally used in AlexNet for distributing computation across GPUs, are repurposed here as a principled mechanism for increasing representational power. Unlike network compression techniques that decompose weight matrices to reduce parameters, ResNeXt focuses on improving accuracy rather than compression. And unlike model ensembling, all branches in ResNeXt are trained jointly as a single model, sharing parameters through the same optimization process.

A standard ResNet block performs a transformation F(x) = w_2 * ReLU(w_1 * x) with a residual connection y = x + F(x). ResNeXt generalizes this by aggregating C identical transformations: F(x) = sum_{i=1}^{C} T_i(x), where C is the cardinality and each T_i has the same topology but separate parameters. In the concrete design, each transformation path uses a bottleneck of 1x1 conv (reducing to d dimensions) -> 3x3 conv -> 1x1 conv (restoring dimensions). The notation "32x4d" denotes 32 paths, each with a 4-dimensional bottleneck. Two simple rules constrain the design: (1) blocks sharing spatial resolution share hyperparameters; (2) when spatial maps downsample by 2x, width doubles.

The aggregated transformations can be equivalently reformulated in three ways: (a) the explicit aggregation form where paths are summed; (b) a concatenation form where low-dimensional outputs are concatenated then processed by a single wider 1x1 convolution; and (c) a grouped convolution form where a single wide layer's channels are divided into groups. All three produce mathematically identical results. The grouped convolution interpretation is most significant for implementation: "all the low-dimensional embeddings can be replaced by a single, wider layer. Splitting is essentially done by the grouped convolutional layer when it divides its input channels into groups." This makes ResNeXt efficiently implementable using standard deep learning frameworks.

On ImageNet-1K, ResNeXt-50 (32x4d) achieves 22.2% top-1 error versus ResNet-50's 23.9%, a 1.7% improvement under the same complexity. ResNeXt-101 (32x4d) reaches 21.2% top-1 error compared to ResNet-101's 22.0%. Crucially, when doubling computational complexity from the ResNet-101 baseline: going deeper (ResNet-200) yields only 0.3% improvement, going wider yields 0.7% improvement, but increasing cardinality (ResNeXt-101 64x4d) yields 1.3% improvement. On ImageNet-5K, ResNeXt-101 shows 2.3% improvement over ResNet-101 on 5K-way classification. On COCO detection, ResNeXt provides consistent improvements of +1.0% AP.

ResNeXt demonstrates that cardinality is a concrete, measurable dimension that is orthogonal to network depth and width. The modularized design with homogeneous transformations proves simpler than Inception variants while achieving superior accuracy on image classification, object detection, and smaller-scale recognition tasks. The architecture generalizes effectively from ImageNet-1K to ImageNet-5K, CIFAR, and COCO, suggesting that the benefits of increased cardinality are robust across scales and domains.