Amap bounds long-horizon 3D mapping to a 12-frame window instead of a growing keyframe store

Twelve frames is eleven cached neighbours plus the one you are working on ^[16]. The repository states the consequence plainly: the model's state and its per-frame work do not depend on how long the sequence has run ^[5]. That is what lets you size a memory budget from a board spec rather than from a route length, which is the opposite of how systems retaining historical anchors, keyframes or feature caches behave as a sequence stretches into thousands of frames ^[6].

That tradeoff shows up as drift over a long sequence. Every estimate is local: geometry for the current frame, pose relative to the frame before it ^[3]. Chain enough of those and rotation error compounds. Amap's answer lives in training rather than in stored history, with a motion-visual rotation refiner and a composition-aware pose loss both documented in the README as drift limiters ^[7]. Revisits are handled by a separate optional backend that retrieves candidate frame pairs and runs sparse pose-graph refinement ^[8]. It sits outside the streaming loop, and the headline Oxford Spires camera-pose result is labelled as streaming-only, without loop closure ^[9]. Labelling a result by which components produced it is good craft, and rarer than it should be.

Ten thousand frames at 24.45 FPS ^[10] is 409 seconds, close to seven minutes of video ^[20]. That is not a long shift for a delivery robot, so the durable part of the claim is the flat per-frame cost rather than the frame count ^[5][19].

For 24.45 FPS to mean anything on the consumer card named in the release, the conditions of the measurement would have to survive the move. The input would still be 504 x 280, which is 141,120 pixels per frame ^[17][11]. The software environment, validated on an A100 ^[11], would have to run on much older silicon. And input storage would have to stay outside the budget, as it was in the H100 measurement ^[11]. A figure measured with input storage excluded describes only the model's compute cost, separate from the robot's actual memory budget. Amap's release leans the other way, using the 6.71 GB peak to argue a GTX 1080 Ti can run the full pipeline ^[10]; runtimewire, reading the repository's benchmark documentation, notes the disclosed evidence does not establish that speed on the older consumer card ^[12].

The memory claim survives better. 6.71 GiB against the 1080 Ti's 11 GB is about 61 percent of capacity, roughly 4.3 GB of headroom ^[18], and the reporting allows that it may fit while asking for a device-specific benchmark before the throughput number travels ^[13].

Adoption cost is the other half. The code is open and the weights are noncommercial ^[14], so a commercial robot programme gets an architecture to reimplement rather than a checkpoint to ship. Credit goes to AMAP CV Lab collectively, with no named project lead and no completed paper citation in the reports ^[15], which leaves the drift-control components resting on README description.

In my context, where a board fixes the memory budget and nobody fixes the route length, the bounded-context bet is the right trade. 24.45 FPS is the first number I would re-measure on my own hardware.