In this paper, we study the effectiveness of features from Convolutional Neural Networks (CNN) for predicting the ambient temperature as well as the time of the year in an outdoor scene. We follow the benchmark provided by Glasner et al.  one of whose findings was that simple handcrafted features are better than the deep features (from fully connected layers) for temperature prediction. As in their work, we use the VGG-16 architecture for our CNNs, pretrained for classification on ImageNet. Our main findings on the temperature prediction task are as follows. (i) The pooling layers provide better features than the fully connected layers. (ii) The quality of the features improves little with fine-tuning of the CNN on training data. (iii) Our best setup significantly improves over the results from Glasner et al. showing that the deep features are successful in turning a camera into a crude temperature sensor. Moreover, we validate our findings also for time prediction and achieve accurate season, month, week, time of the day, and hour prediction.