TPose: Reformulating Thin-Object Detection as Top-Down Pose Estimation

Thin objects such as road lane lines and other elongated structures are challenging to detect and localize because they occupy a small fraction of an image, exhibit slender and non-compact geometry, and often occur against noisy backgrounds. Bounding-box representations provide inaccurate localization for such objects, whereas pixel-wise representations introduce severe foreground-background imbalance, computational overhead, and the need for additional post-processing to recover object geometry. We investigate an alternative formulation in which thin-object detection is treated as top-down pose estimation, representing each object using a small set of semantic keypoints. We introduce TPose, a generic framework for thin-object detection comprising three components: (1) an annotation module that identifies an appropriate keypoint representation for a target thin object; (2) kMobileNetV3, a lightweight extension of MobileNetV3 with a three-layer deconvolutional head for keypoint-based detection; and (3) an Unbiased Data Processing (UDP) module that improves semantic alignment between input and predicted keypoints during transformation and inference. By representing thin objects through keypoints, TPose avoids the localization limitations of coarse bounding boxes and the dense prediction burden associated with pixel-wise representations. Furthermore, because the task is formulated as pose estimation, the relationships between detected keypoints directly define the thin object, eliminating the need for explicit post-processing to reconstruct its geometry. We evaluate the transferability of this formulation to road lane-line detection through RLPose, an instantiation of TPose using three representative keypoints per lane line: start, center, and end. We develop a corresponding annotation protocol and evaluate RLPose on the TuSimple benchmark, containing 358 training and 100 test images at 256 × 128 resolution. The proposed kMobileNetV3-Large achieves 64.1 ± 0.75% AP, 65.36 ± 0.70% AR, and 64.7 ± 0.70% F1-score, outperforming the evaluated HRNet-W32, HRNet-W32+UDP, and ResNet-50 Simple Baseline pose-estimation networks on these primary metrics. It achieves these results with only 5.23M parameters and an inference speed of 29.01 FPS, compared with 28.54M parameters and 7.08 FPS for HRNet-W32 and 34M parameters and 32.41 FPS for ResNet-50. These results provide empirical evidence that thin-object detection can be effectively reformulated as pose estimation and that a common keypoint-based framework can transfer to a distinct thin-object domain without requiring an application-specific detection architecture. The findings support keypoint-based representation as a promising direction for accurate and computationally efficient thin-object detection. The current three-keypoint representation remains less effective for highly curved lane lines, motivating future work on adaptive keypoint representations for complex thin-object geometries.