CMSNet: Modular Segmentation for Off-Road Use

• CMSNet is a configurable modular framework designed for off-road semantic segmentation, focusing on obstacle and passable-region detection in challenging environments.
• It integrates interchangeable backbones, pyramid modules, and adjustable output strides to balance accuracy and latency for real-time deployment.
• Field tests and ablation studies on the Kamino dataset demonstrate CMSNet's robust performance with up to 86.98% mIoU under various adverse conditions.

The Configurable Modular Segmentation Network (CMSNet) is a semantic-segmentation framework for autonomous vehicles and advanced driver assistance on unpaved roads and off-road environments, particularly under adverse visibility and low-latency constraints. It was introduced to support segmentation of obstacles and trafficable ground in scenes where the vehicle may need to navigate rough terrain without a predefined trail, and it was developed together with the Kamino dataset and a real-time deployment pipeline based on TensorRT, C++, and CUDA (Alves et al., 2020, Neto, 20 Sep 2025).

1. Operational context and task formulation

CMSNet was developed for perception in unpaved-road and off-road environments, including non-uniform terrain in open-pit mines and developing countries. The target problem is not standard lane-based road parsing. Instead, the system addresses detection of regions where the vehicle can pass through when there is no clear boundary of the track, while also segmenting obstacles under night, rain, dust, low light, and other visibility impairments. The reported application domain therefore combines semantic scene understanding with hard real-time constraints and domain shift relative to urban-road benchmarks (Alves et al., 2020).

The central segmentation objective is to distinguish obstacles from areas where the car can pass through. The associated evaluation program includes a study of how the proposed segmentation algorithms behave in different severity levels of visibility impairment and field tests carried out with semantic-segmentation architectures ported for real-time inference. A plausible implication is that CMSNet was conceived less as a single fixed model than as an engineering and experimental framework for selecting architecture–latency trade-offs in safety-critical off-road perception.

2. Modular architecture and configurable design

CMSNet is described as a flexible and configurable framework that enables different architectural arrangements. Its modularity is expressed through interchangeable architectural elements rather than through a single immutable network definition. The configurable dimensions reported for CMSNet are backbone selection, pyramid module, output stride, shortcut usage, and the decoder’s upsampling strategy (Alves et al., 2020).

Component	Reported options	Reported role
Backbone	MobileNetV2, ResNet, VGG	Feature extraction
Pyramid module	SPP, ASPP, GPP	Multi-scale or global context aggregation
Output stride	OS8, OS16	Accuracy–latency trade-off
Shortcut connections	Enabled or disabled	Fine-detail reconstruction during upsampling
Upsampling	Convolution + bilinear interpolation	Efficient decoding

MobileNetV2 is identified as the default backbone for efficiency, and the paper summary states that it was slightly modified by removing later convolution and pooling layers, reducing parameters from 3.5M to 1.84M. The pyramid stage can be instantiated as Spatial Pyramid Pooling, Atrous Spatial Pyramid Pooling, or Global Pyramid Pooling. Output stride is configurable as OS8 or OS16, and shortcut connections can be turned on or off. Instead of transposed convolution, the decoder uses standard convolution followed by efficient bilinear interpolation for upsampling (Alves et al., 2020).

The framework was used to instantiate a family of variants denoted CM0–CM8. The accompanying description emphasizes that the main novelty lies in the configurable, modular, and efficiently implemented integration of established modules such as SPP, ASPP, GPP, and skip connections rather than in the introduction of a new primitive segmentation operator. This point is important because CMSNet is sometimes read as though it were a single architecture; the published description instead presents it as a controlled design space for ablation and deployment-oriented selection.

3. Kamino dataset and annotation regime

CMSNet was developed and evaluated together with Kamino, a dataset for unpaved-road and off-road perception. The reported corpus contains almost 12,000 images, with 11,479 annotated images in the more detailed description, and was collected from an operating vehicle with eight synchronized cameras. The dataset includes scenes from an off-road proving ground and real-world unpaved roads, including urban, rural, industrial, and mining environments, with adverse situations such as rain, dust, and low light; the 2025 abstract further describes the proving ground as emulating a mine under adverse visibility (Alves et al., 2020, Neto, 20 Sep 2025).

Kamino is reported to have a high number of labeled pixels compared to similar public collections. The annotations are described as panoptic, so both semantic- and instance-segmentation ground truth are available, although the reported CMSNet study evaluates semantic segmentation. The class inventory is summarized as road, person, car, truck, bike, bus, animal, cone, plus background. A reduced split, Kamino-Small, contains 5,523 images for faster iteration, with the remainder used for validation (Alves et al., 2020).

The dataset design is tightly coupled to the CMSNet research questions. It contains adverse conditions such as night, rain, dust, muddy or dirt roads, and also synthetic impairments including fog and Gaussian noise. Off-road sections were deliberately constructed to lack clear visual boundaries, for example where a sand road blends with the background. This suggests that Kamino was intended not merely as a benchmark extension but as an instrument for probing failure modes of semantic segmentation when the notion of “drivable region” is weakly delimited by appearance alone.

4. Evaluation protocol, ablations, and reported performance

The reported evaluation metrics are Pixel Accuracy ($P_{acc}$), Classwise Accuracy ($CP_{acc}$), Mean Accuracy ($mCP_{acc}$), Mean Intersection-over-Union ($mIoU$), Frequency Weighted IoU ($FWIoU$), and inference time in frames per second with standard deviation. Extensive ablation studies were conducted across nine CMSNet variants, with the framework enabling fair comparison by varying single modules while keeping most of the pipeline fixed (Alves et al., 2020).

On Kamino, the best reported configuration reaches 86.98% $mIoU$, identified as CM2 with OS8 and ASPP. The same summary reports class-wise IoU values of up to 96.5% for Road, 78.7% for Car, and 75.9% for Person. CMSNet variants are described as outperforming off-the-shelf DeepLab and PSPNet models trained on urban datasets, which are reported at 47.99%–57.83% $mIoU$ on this off-road task, and CMSNet models are reported at 16–32 FPS on RTX2060-class GPUs compared with 0.7–2.7 FPS for the cited DeepLab and PSPNet baselines (Alves et al., 2020).

The ablation results are summarized in qualitative architectural terms. Lower output stride, specifically OS8, slightly improves accuracy but slows inference. ASPP generally yields the best $mIoU$ but is heavier than the lighter context modules. On the DeepScene dataset, CMSNet is described as outperforming prior off-road-specific networks such as cnns-fcn and dark-fcn, while matching or slightly edging UpNet. The two-dataset validation claim in the 2025 abstract is consistent with this broader experimental framing (Alves et al., 2020, Neto, 20 Sep 2025).

Robustness analysis under adverse visibility is a central part of the empirical picture. The summary reports less than 2% $mIoU$ loss in dust, around 5% in night conditions, 8–23% in rain depending on architecture, and larger degradations under severe fog or noise. GPP-based shortcuts are said to help mitigate some performance drop under synthetic fog and noise, whereas shortcuts become less effective as conditions worsen. These observations reinforce the interpretation of CMSNet as a system for controlled trade-off analysis under operational perturbation rather than only a benchmark-optimized model.

5. Real-time optimization and field deployment

Real-time deployment is presented as a first-order design constraint. To achieve embedded real-time inference, CMSNet configurations were ported by removing and fusing many layers using TensorRT, C++, and CUDA. The optimizations reported include model simplification, preference for lightweight modules, layer fusion, memory and compute reduction through bilinear interpolation instead of deconvolution, and use of Global Pyramid Pooling where possible for computational efficiency (Alves et al., 2020).

The reported hardware platforms include NVIDIA Drive PX2 and several GPUs, specifically GTX1050, GTX1060, RTX2060, and GTX 1080TI. After optimization, CM0-TRT is reported at 7.9 FPS on Drive PX2 and CM3-TRT at 21.2 FPS, while CM3-TRT reaches up to 99 FPS on GTX 1080TI. The summary states that CM3-TRT is about $2\times$ faster, with a small drop in accuracy compared to the maximum-accuracy CMSNet variants. At 30 km/h and 21 FPS, information is reported to be available 47 ms after capture, corresponding to 39 cm of vehicle travel (Alves et al., 2020).

Field tests were carried out on an actual vehicle described as a utility van with a multi-camera and LiDAR sensor suite. CMSNet is reported to have achieved robust segmentation of passable road and obstacles in real-world off-road and unpaved scenarios, including night, dust, and rain. A plausible implication is that the framework’s practical value depends as much on deployability and latency predictability as on segmentation accuracy; this is reinforced by the reported low standard deviation of inference time on embedded ARM64 hardware.

6. Relation to modular segmentation research and interpretive cautions

Within the broader segmentation literature, CMSNet belongs to a family of modular approaches that treat segmentation systems as compositional assemblies rather than indivisible monoliths. A relevant earlier example is the hierarchical semantic-segmentation method based on serially connected fully convolutional networks, where one FCN segments a vessel region and a second FCN segments vessel contents using the first module’s output through valve filter attention (Eppel, 2017). That earlier work emphasizes independent training, reuse of trained modules, and improved performance over a single-step network.

This suggests a useful distinction between two notions of modularity. In the hierarchical FCN work, modularity is serial and task-decompositional: one module’s segmentation map becomes a structured input to a downstream module (Eppel, 2017). In CMSNet, modularity is configurational and architecture-compositional: the reported design space is organized around interchangeable backbones, pyramid modules, output strides, and shortcut settings for a single off-road semantic-segmentation pipeline (Alves et al., 2020). The concepts are related, but they are not identical.

A further interpretive caution concerns the documentary record. The 2025 paper “Vision-Based Perception for Autonomous Vehicles in Off-Road Environment Using Deep Learning” reiterates the main system-level description of CMSNet, the Kamino dataset, adverse-condition training, and real-time deployment in its abstract, but the accompanying record explicitly states that it does not provide information regarding architecture, novel techniques, datasets, training setups, performance metrics, empirical results, or real-time inference beyond that high-level description (Neto, 20 Sep 2025). For technical detail, the more granular characterization available in the supplied record is therefore the 2020 CMSNet account (Alves et al., 2020).

Taken together, the published descriptions present CMSNet as a configurable modular framework for off-road semantic segmentation, anchored in a dedicated adverse-visibility dataset and a deployment pathway for embedded real-time perception. Its defining contribution is the integration of established segmentation components into a systematic experimental and deployment framework for environments where drivable-region detection cannot rely on explicit lane or track boundaries (Alves et al., 2020).