Mattia Segu

PhD student at @ETH_en. Member of the Max Planck ETH Center for Learning Systems. Student researcher at @Google.

[Mattia Segu @MattiaSegu]

🎯 I'm excited to introduce SambaMOTR, a novel tracker that models long-range dependencies, tracklet interdependencies, and temporal occlusions in complex scenarios like dance performances, sports, and dynamic animal groups. 🕺⚽️🦅

👾 Project Page: https://t.co/ki92ovx7K3

[1/n]

[Mattia Segu @MattiaSegu]

@carlo_sferrazza Huge congrats Carlo! Excited to hear you are coming back to Zurich

[Mattia Segu @MattiaSegu]

Starting in 5 minutes, see you around :)

[Mattia Segu @MattiaSegu]

🚨 Next week I'll be at #ICLR2025 in Singapore, presenting our Spotlight paper:

Samba: Synchronized Set-of-Sequences Modeling for Multiple Object Tracking 🎉

📍 Fri Apr 25, 10:00–12:30
📍 Hall 3 + Hall 2B, Poster #81
🔗 Project: https://t.co/ki92ovwzUv

Swing by and say hi! 👋

[Mattia Segu @MattiaSegu]

TL;DR: We propose SambaMOTR, a new E2E multiple object tracking framework that learns to track entirely from data using Samba, our novel set-of-sequences model.

[Mattia Segu @MattiaSegu]

🚨 Next week I'll be at #ICLR2025 in Singapore, presenting our Spotlight paper:

Samba: Synchronized Set-of-Sequences Modeling for Multiple Object Tracking 🎉

📍 Fri Apr 25, 10:00–12:30
📍 Hall 3 + Hall 2B, Poster #81
🔗 Project: https://t.co/ki92ovwzUv

Swing by and say hi! 👋

[Mattia Segu @MattiaSegu]

@wightmanr That is awesome, thank you so much @wightmanr! I'll be trying the new weights out soon

[Mattia Segu @MattiaSegu]

@wightmanr @wightmanr I can confirm that the ImageNet-12k-pretrained model is noticeably better than the standard one also for the hybrid medium model

[Mattia Segu @MattiaSegu]

@wightmanr Awesome! I can launch the experiments on the hybrid medium tomorrow and let you know soon, the gain should be visible already from the first epochs

[Mattia Segu @MattiaSegu]

@wightmanr it seems that the huge performance boost (+6 mAP on COCO) was due to ImageNet-12k pretraining. The anti-aliasing version without ImageNet-12k pretraining performs slightly worse than the conv large without both ImageNet-12k and antialising.

Interestingly, I don’t see this difference for FasterViT, where ImageNet-12k pretraining does not help.

Do you think that it would be possible to have the other MobileNet-v4 models with ImageNet-12k pretraining?

[Mattia Segu @MattiaSegu]

Thanks for the reply! I’m now training starting from the equivalent conv large aa model without ImageNet-12k pretraining to double check whether it’s really aa that makes the difference. I’ll keep you posted

I’m definitely looking for the smaller-sized models with anti-aliasing, but it would be interesting to have also the hybrid medium and large if it isn’t asking too much

[Mattia Segu @MattiaSegu]

@lekkima @wightmanr I think it comes from anti-aliased CNNs https://t.co/tZcDxSJfOq

it is not discussed in the MobileNet-V4 paper, it was added to the timm implementation: https://t.co/xiF65CgxLa

[Mattia Segu @MattiaSegu]

@wightmanr Perhaps the boost also comes from the ImageNet21k pretraining. Either way, it would be nice to have that for all MobileNetv4 models :)

[Mattia Segu @MattiaSegu]

Hey @wightmanr, great job integrating MobileNet-V4 in timm! I'm finding that the anti-aliased conv large version works so much better as initialization for object detection compared to the standard conv large version. Any chance that you would also provide the anti-aliased weights for the other model sizes?

[Mattia Segu @MattiaSegu]

This work wouldn't have been possible without my collaborators: @lpiccinelli944, @derek_siyuanli, @YungHsuYang, Bernt Schiele, and Luc Van Gool.

👾 Project Page: https://t.co/ki92ovx7K3
📕 Paper: https://t.co/XfG2GypN1g

[Mattia Segu @MattiaSegu]

At the heart of SambaMOTR is Samba, a set of synchronized state-space models. Each tracklet’s memory is updated based on observations, then synchronized. This ensures that long-term dependencies and interactions among tracklets are captured to predict the next states accurately.

[Mattia Segu @MattiaSegu]

SambaMOTR excels in complex datasets like DanceTrack (synchronized dance), SportsMOT (interdependent player movements), and BFT (bird flock tracking). The result? Accurate, robust tracking despite occlusions and complex dynamics. 💥

[Mattia Segu @MattiaSegu]

SambaMOTR combines a transformer-based object detector with our novel set-of-sequences Samba model in a tracking-by-propagation framework. Samba jointly processes the tracklets’ memory and their current observations to predict future queries and update track memory.

[Mattia Segu @MattiaSegu]

🎯 I'm excited to introduce SambaMOTR, a novel tracker that models long-range dependencies, tracklet interdependencies, and temporal occlusions in complex scenarios like dance performances, sports, and dynamic animal groups. 🕺⚽️🦅

👾 Project Page: https://t.co/ki92ovx7K3

[1/n]

[Mattia Segu @MattiaSegu]

Walker wasn't our only tracking paper at #ECCV2024! 🔍

@derek_siyuanli and @YungHsuYang presented SLAck, an open-vocabulary tracker merging semantic, location, and appearance cues to boost association. 🎯

Check it out: https://t.co/TJttqAKsE6

That's a wrap from Milan! 👏

[Mattia Segu @MattiaSegu]

@sidgairo18 I had to choose between a picture with the side-eyed look and a picture without the crowd 😅 don’t make me regret my choice 😂

[Mattia Segu @MattiaSegu]

🎉🚶‍♂️ Thrilled to have presented our self-supervised multiple-object tracker Walker at #ECCV2024 in my home country! 🚶‍♂️🎉

Big thanks to everyone who stopped by for all the great discussions! 🙌

Learn more about our paper here: https://t.co/KyMkY3MySp