Supplementary Material: Learning Representations from Audio-Visual Spatial Alignment

These are transformer networks of base dimension 512 and expansion ration 4. In other words, All models were trained using the Adam optimized. Pre-training hyper-parameters are summarized in Table 2. For semantic segmentation, we used a lightweight FPN segmentation head. Semantic segmentation predictions are then computed based on the features at all levels. This shows the use of spatial negatives is complementary to AVC.