Nate@nate_512

Follow

PICK UP A CUP. SOUNDS SIMPLE? NOW LABEL EVERY MOVE. Reach. Grasp. Lift. Hold. Put it down. a single label for the whole clip misses that detail! 🚨 Perceptroninc’s new Egocentric system breaks video into individual steps, with a separate action-label timeline for each hand or gripper. The app is built with Svelte and uses the self-service APIs to connect the entire process: > upload videos > follow annotation jobs > download the labels The team’s latest numbers are just INSANE: → 12–25% better hand-shape reconstruction accuracy than ACE-Ego-Hand on HOT3D and ARCTIC → 0.583 segment F1 on WGO-Bench, described as the highest reported score → EgoHome, a new benchmark focused on subtasks shorter than a second It also supports up to 1,000 videos per batch 🔥 For robotics teams, that combination matters: detailed annotations for individual hands, with a workflow built to handle volume. Full results in 🧵 ↓

1w

Nate@nate_512

Follow

PICK UP A CUP. SOUNDS SIMPLE? NOW LABEL EVERY MOVE. Reach. Grasp. Lift. Hold. Put it down. a single label for the whole clip misses that detail! 🚨 Perceptroninc’s new Egocentric system breaks video into individual steps, with a separate action-label timeline for each hand or gripper. The app is built with Svelte and uses the self-service APIs to connect the entire process: > upload videos > follow annotation jobs > download the labels The team’s latest numbers are just INSANE: → 12–25% better hand-shape reconstruction accuracy than ACE-Ego-Hand on HOT3D and ARCTIC → 0.583 segment F1 on WGO-Bench, described as the highest reported score → EgoHome, a new benchmark focused on subtasks shorter than a second It also supports up to 1,000 videos per batch 🔥 For robotics teams, that combination matters: detailed annotations for individual hands, with a workflow built to handle volume. Full results in 🧵 ↓

1w