【論文紹介】Attention-GAN

Attenotion-GAN for Object Transfiguration in Wild Images

https://arxiv.org/abs/1803.06798

Transcript

1. ATTENTION-GAN FOR OBJECT TRANSFIGURATION IN WILD IMAGES ʲ࿦จ঺հʳ
2. ֓ཁ Attention-Gan For Object Transfiguration In Wild Images ▸ ॻࢽ৘ใ https://arxiv.org/abs/1803.06798 Submitted : 19 Mar 2018 ▸ ֓ཁ ը૾υϝΠϯؒͷࣸ૾ʢม׵ʣΛֶश͢ΔωοτϫʔΫ AttentionػߏΛऔΓೖΕΔ͜ͱʹΑͬͯ஫໨͢΂͖ྖҬ͚ͩࣸ૾͢Δ͜ͱ͕Ͱ͖ɺ ΑΓ៉ྷͳը૾Λੜ੒͢Δ͜ͱʹ੒ޭ
3. Ϟσϧ Attention-GAN ஫໨ྖҬΛ੾Γग़͢Attention Network + ม׵Λ୲͏Transformation Network
4. Ϟσϧ Attention GAN = Cycle GAN + Attention Mechanism ▸ Cycle consistency loss ▸ Attention loss X →Y →X Y→X →Y Attention cycle-consistent loss Attention sparse loss ஫໨ྖҬ͕ՄೳͳݶΓখ͞ͳʢεύʔεͳʣྖҬʹͳΔΑ͏ʹL1ϊϧϜΛ௥Ճ
5. Ϟσϧ Supervised Learning ▸ Attention ≒ Segmentation Segmentation label͕͋ΔͷͳΒAttention network͸
 segmentation໰୊Λղ͘Α͏ʹֶशͤ͞Ε͹ྑ͍ Attention supervised loss Total loss λ͸cycle consistent loss ͱ attention loss ͷॏཁ౓Λίϯτϩʔϧ͢Δ
6. ֶश Experiments ▸ Datasets ImageNet: tigert 1444 images, leopard 1396 images MSCOCO: horse, zebra 286 x 286 ʹϦαΠζͨ͠ޙɺϥϯμϜʹ256 x 256Ͱ੾Γग़͠ ▸ Training strategy Optimizer: Adam, LR: 0.0002 (~100epoch), 0ʹͳΔΑ͏ʹઢܗʹݮਰ(~200epoch) Batch size: 1
7. ݁Ռ Results
8. ݁Ռ Comparison with CycleGAN ▸ ఆੑൺֱʢࠨஈ: input, தஈ: AttentionGAN, ӈஈ: CycleGAN) ▸ ఆྔൺֱ AttentionGANͰ͸ ಛʹinputͷഎܠ৘ใΛอ͍ͯͯΔ AttentionGAN, CycleGANͷ ͲͪΒ͕ΑΓྑ͍ը૾Λ ੜ੒͍ͯ͠Δ͔ͷΞϯέʔτ݁Ռ (੨: AttentionGAN, ੺: CycleGAN)
9. ߟ࡯ Ablation Analysis ▸ AttentionͷॏΈΛมߋ Sparse lossͷ܎਺Λมߋ͍ͤͯ͘͞ͱ… ࣮ݧతʹ λ=1 ͕ϕετͩͬͨ
10. ߟ࡯ Comparison of Supervised Results segmentation label ࠐΈͰֶश͢Δํ͕ྑ͍݁Ռʹ
11. ·ͱΊ Conclusion ▸ GANʹAttentionػߏΛ૊ΈࠐΉࣄʹΑΓɺΑΓϦΞϧͳυϝΠ ϯؒࣸ૾ΛՄೳʹͨ͠ ಛʹഎܠ෦ͷྼԽΛ཈͑Δ͜ͱʹ੒ޭ ▸ ஫໨ྖҬʹΑΓ஫ҙΛ෷͏Α͏ʹɺattention cycle consistent- loss, attention sparse lossΛఏҊ