Geoguessr

Transcript

1. Geoguessr 박상언
2. geoguessr.com
3. Guessing • 태국어 -> 높은 확률로 태국 • (태국어 해석하면 위치도 알 수 있겠지만 귀찮음) • 방콕 정도로 예측
4. Result
5. geosettr.com
6. ?! • http://arxiv.org/abs/1602.05314
7. • Input: Image pixels • Training: CNN • Output: One-hot encoded vector of map shown above
8. • 10 well-traveled human subjects VS PlaNet • No pan, zoom • PlaNet wins 28/50 rounds
10. Sequence Geolocation with LSTM • Embedding vector from final layer before softmax • Feed vector to LSTM • Output of LSTM to softmax layer • Use in image sequences (like albums)
11. • PlaNet avg: Average on album • LSTM off: Using offsets (View several pictures before inference) • LSTM rep: Repeated Sequences, but predictions from the end of the sequence get carried over to the beginning. • BLSTM: Bidirectional LSTM works good, but may be too slow for long sequences