Soft Hough Forest-ERTs: Generalized Hough Transform based object detection from soft-labelled training data

• A modified soft label estimation method by selecting a reliable positive bag based on Maximum Mean Discrepancy.
• Extremely Randomized Trees are extended to learn from soft-labelled training blobs.
• Probabilistic Hough voting process is derived from soft label ERTs codebook.
• Weakly supervised object detection method is proposed.
• Experimental results show the advantage of utilizing soft labels, and the performance of the proposed weakly supervised object detection method.

Classical supervised object detection methods learn object models from labelled training data. This is tedious to create especially when the training dataset is large. Detection methods such as background subtraction and headlight detection can detect potential positive blobs that may contain the object without labelled training data. However, such blobs are not always accurate. They may include noise such as part of an object, multiple objects and other types of objects. Therefore, soft labels that indicate their probability of being positive may be more useful. A modified soft label estimation method based on Maximum Mean Discrepancy is introduced in this work. Based on it, a Generalized Hough Transform based object detection method from soft-labelled training data is proposed to utilize potential detections and their estimated soft labels. Experimental results show that the method can achieve comparable performance to supervised methods. It outperforms both Generalized Hough Transform based object detection with hard-labelled training blobs, and a state-of-the-art weakly supervised method.