Although the new vibration sensor board will still take a week.

The Algorithm Group wasn't idling either.

They pushed through all possible software-level optimizations for gain compensation.

Currently, the missed detection rate for soft surfaces has been squeezed down from thirty-seven percent to twenty-four percent.

Just like what he said to Chen Wen before.

Going any further is really impossible to push down.

It can be said that the software gain has reached its ceiling now; going any higher would mean the noise floor drowns out the signal, and the signal-to-noise ratio would instead only deteriorate.

"With the new hardware not in place, this optimization can only be stuck here."

The Algorithm Group Leader said to Xiao Zheng.

"Let's switch to something else."

Subsequently, he opened the training set management panel and glanced at the data distribution.

By now, the joint vibration annotation framework had been running for almost a week.

The samples automatically entered into the set had risen from twenty-three to over one hundred seventy.

Coupled with manually confirmed ones, the total volume of the training set expanded from the original three-hundred-plus to close to six hundred.

Then, he let the model run another round on this batch of new data.

The boundary sample confusion rate dropped from the previous nineteen percent to the current fourteen percent.

"It's still dropping, but the speed is getting slower and slower."

The Data Team Leader pointed at the loss curve and said.

"Especially after the sixth round, it basically leveled off."

It's the same problem as last time: the amount of data has increased, but the quality hasn't kept up.

Although most of the newly admitted samples are 'clean', the truly difficult boundary cases remain the same.

The Algorithm Group Leader flipped through the annotation records.

Among the more than one hundred seventy newly entered items, over eighty percent were "high-confidence" samples.

Among them were those with large vibration amplitudes, clear sounds, and clear IMU data.

Because the model had long since learned these samples, feeding it more afterwards would yield very limited improvements.

What truly stalled the confusion rate was actually just that small handful of tough nuts.

For example, multiple people in the same frame, soft surfaces, complete stillness and silence, and so on.

The Algorithm Group Leader said.

"We still have to smash into the difficult samples."

Xiao Zheng asked.

"How do we smash into them?"

"Active mining."

From those millions of hours of data, specifically filter out clips of these scenarios.

Manually fine-annotate them, and then feed them to the model for targeted training."

Subsequently, he brought up the data retrieval conditions and entered a few keywords:

Multiple people in the same frame, carpet/mattress, no crying, stillness exceeding five seconds after falling, and so on.

The retrieval results popped out: over two hundred fifty-four thousand three hundred candidate clips.

"This is just a rough keyword filter."

The actually matching ones might only be around one hundred seventy thousand items."

"So many, will all of them be manually fine-annotated?"

"Of course… impossible. With such a huge workload, how much manpower and material resources would that cost?"

"We don't need to annotate all of them."

"We only need to pick some out from these and then feed them to the model."

"But these cannot be like the previous 'yes or no' confirmations."

Instead, it's frame-by-frame annotation.

Annotate every person's skeleton keypoints, ground contact status, and object occlusion relationships in every frame."

"I calculated it, it'll probably be just five or six hundred items."

Xiao Zheng also calculated it and said.

"A thirty-second video takes about an hour and a half to annotate frame-by-frame."

Five hundred items would be seven hundred fifty hours of annotation work hours.

And this is still the minimum."

"With the six people in our group, throwing everyone at it will take at least two weeks."

"Then two weeks it is."

While speaking, the Algorithm Group Leader was already creating the task list.

"Let the annotation group run first."

I will synchronously do data augmentation on my side, performing things like rotation, cropping, lighting changes, and occlusion simulation on the well-annotated samples.

Turn one into ten to expand the training set."

He paused for a moment and added another item.

"In addition, for samples with multiple people in the same frame, the annotation rules need to be refined."

Currently, our annotations only mark the 'main target'."

That is, the largest person in the frame or the person closest to the robot.

But in home scenarios, two people are often stacked on top of each other."

For instance, an adult holding a child and falling, a child being pinned underneath an adult, two children crashing into each other, and so on."

Hearing this, Xiao Zheng asked again.

"Who should be annotated in this situation?"

"Annotate both."

Each person has an independent annotation trajectory, while simultaneously marking the spatial relationship between them."

Such as contact, occlusion, and separation."

The model must learn to distinguish between 'a child falling by themselves' and 'an adult holding a child falling together'."

The early warning levels for these two situations should be different."

In the latter, an adult is present, so the risk is lower; in the former, the child is alone, so the risk is higher."

Hearing this, the Data Team Leader also nodded and said.

"The annotation complexity multiplies several times over like this."

"But it's worth it."

This is the highest-frequency boundary scenario in real homes."

Xiao Zheng had already started pulling up the annotation task list.

The Algorithm Group Leader walked over to the whiteboard and added a line below "Multimodal Weakly Supervised Framework".

"Second round of training: targeted mining of boundary scenarios, multi-person multi-target annotation, data augmentation."

Goal: suppress the confusion rate from 14% to below 8%."

After writing, he glanced at the adjacent "Vibration Sensor Hardware Gain" line which was still marked as "Waiting", with an estimated four days remaining.

Four days.

He could use these four days to finish running the first round of boundary sample annotations, and directly conduct joint verification once the board arrives.

Then he opened the first retrieved video.

In the footage, a child over two years old rolled off the sofa cushion, landed in the middle of a pile of plush toys, and was completely blocked from view.

Crawling out three seconds later, completely fine.

Xiao Zheng watched a few frames and spoke.

"Who should this one be annotated as? The child is blocked by toys, and half of the skeleton keypoints cannot be extracted."

"Then annotate the parts that can be extracted, and mark the unextractable parts as 'occluded'."

The model must learn to handle incomplete information; after all, there will never be a shortage of occlusions in real scenarios."

"Then how is the confidence calculated?"

"For frames where occlusion exceeds fifty percent, that signal channel is marked as 'unavailable', and its weight is automatically lowered during joint judgment."

You can't guess blindly just because you can't see, nor can you directly discard the entire piece of data just because you can't see."

Hearing this, Xiao Zheng nodded and began drawing bounding boxes for annotation.

Subsequently, the Algorithm Group Leader returned to his screen and opened the data augmentation script.

Rotation, scaling, brightness perturbation, random occlusion, and so on—he set the parameters well and ran a test.

The augmented samples looked like different angle and lighting versions of the same scenario, and the skeleton keypoint coordinates were automatically transformed.

"It can run."

He said to the Data Team Leader.

"Keep an eye on the annotation group; raise the random inspection ratio of annotation quality to twenty percent."

Boundary samples cannot have errors; if even one is wrong, the model will learn crookedly."

"Understood."

Suddenly, a message popped up in the bottom right corner of the screen.

It was sent by Chen Wen.

"The PCB redesign layout is done; it will be sent to the factory for prototyping tomorrow."

Expected back in three days."

One day faster than previously estimated."

He replied with "Okay", and then casually changed "Waiting" on the whiteboard to "Prototyping, estimated 3 days".

Then he glanced at the confusion rate curve again: fourteen percent.

There is still some distance from the target of eight percent, but the direction is correct, so there shouldn't be much of a problem.

Xiao Zheng's voice came from the annotation group's side.

"Boss, in this video the child fell, but there is also a cat running past from the side in the frame. Does the cat count as an annotation target?"

"Counts. Directly mark it as 'other dynamic objects', do not participate in human skeleton extraction, but mark its position information."

Because the model also needs to learn to distinguish between scenarios like 'a child falling' and 'a cat running past'."

"Okay."