Smoke detection is crucial for preventing fires. However, existing object detectors struggle to detect smoke due to its variable morphology, fluidity, and transparency compared to general objects. This paper identifies three challenges in smoke detection: the limited availability of public smoke datasets, the inability of normal convolution and standard bounding box encoding schemes in locating smoke, and the adverse effect of ordinary parallel detection heads on detection accuracy. To address these issues, we first construct a smoke dataset, IndoorSmoke9000, and propose a new network for smoke detection, called RepPoints++, which applies deformable convolution (DCN) in feature extraction to adapt to smoke’s dynamic shape and more accurately determine the key points in the target. We then apply a soft key points encoding scheme (Soft-Min-Max) to address positioning inaccuracies. Finally, we introduce a shared convolution head (SC-Head) to enhance interaction between the classification and regression branches, improving detection accuracy. Experiments conducted on our IndoorSmoke9000 dataset, an open-source large multi-class object dataset (the COCO dataset), and an open-source large single-class outdoor smoke dataset (the Outdoor dataset) demonstrate that RepPoints++ not only achieves state-of-the-art performance on both the IndoorSmoke9000 and Outdoor datasets, but also shows the generalization on the COCO dataset.