good work!

[Daniil-Nay]

Thank you for your hard work!
However, I noticed your advice in a closed issue thread about searching for a new tracker.

Sorry for the slightly off-topic question, but are you by any chance aware of any new lightweight trackers? It seems quite challenging for me to find lightweight transformer-based trackers (I’ve gone through your article and some benchmarks but haven’t found any except for HiT and FeAR).

[TsingWei]

The newest I know is SUTrack(AAAI 2024). He provide a tracker family from tiny to heavy, by applying fast-iTPN family as backbone.

[alaa-i]

I would suggest AVTrack
AVTrack

[TsingWei]

Basically u need a lightweight backbone to get a efficient tracker. If the light-est backbone u could find cannot satisfy the computation constrain of your device, u can try some techniques to further accelerate it.

@alaa-i In practical applications, we prefer the tracker to have a fixed runtime latency. However, AVtrack and other algorithms with dynamic early-exit mechanisms (which are common not only in tracking tasks but also in other computer vision tasks) can only achieve better performance in terms of average runtime latency. They cannot guarantee an upper bound on the runtime latency for every frame.

[Daniil-Nay]

thanks a lot! Appreciate it.

[TsingWei]

By the way I'm working on developing extremely tiny trackers designed for low-cost edge devices like Raspberry Pi. Stay tuned if you're interested!

[alaa-i]

By the way I'm working on developing extremely tiny trackers designed for low-cost edge devices like Raspberry Pi. Stay tuned if you're interested!

Nice to know !
I sure hope to see it soon. Is there an expected release date ?

[sam-s22]

By the way I'm working on developing extremely tiny trackers designed for low-cost edge devices like Raspberry Pi. Stay tuned if you're interested!

Any news regarding your new trackers ?