Music emotion recognition (MER) studies have made great progress in detecting the emotions of music segments and analyzing the emotional dynamics of songs. The overall emotion and depth information of entire songs may be more suitable for real-life applications in certain scenarios. This study focuses on recognizing the overall emotion and depth of entire songs. First, we constructed a public dataset containing 3839 popular songs in China (PSIC3839) by conducting an online experiment to collect the arousal, valence, and depth annotation of each song. Second, we used handcrafted feature-based method to predict the overall emotion and depth values. Support vector regressions using Mel frequency cepstrum coefficients features as inputs achieve good model performance (arousal: R² = 0.609; valence: R² = 0.354; and depth: R² = 0.465). Finally, the groupwise and personalized results were also investigated by training a unique regressor for each group or individual, which provides a reference for future research.